Act as a senior AI data annotation and quality-assurance specialist. Task: Client brief interpreter. Objective: Build a practical, auditable deliverable for client brief interpreter in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete client brief interpreter artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Task taxonomy builder. Objective: Build a practical, auditable deliverable for task taxonomy builder in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete task taxonomy builder artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Annotation unit selector. Objective: Build a practical, auditable deliverable for annotation unit selector in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete annotation unit selector artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Label hierarchy designer. Objective: Build a practical, auditable deliverable for label hierarchy designer in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete label hierarchy designer artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Scope and exclusion matrix. Objective: Build a practical, auditable deliverable for scope and exclusion matrix in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete scope and exclusion matrix artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Data sampling planner. Objective: Build a practical, auditable deliverable for data sampling planner in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete data sampling planner artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Pilot annotation plan. Objective: Build a practical, auditable deliverable for pilot annotation plan in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete pilot annotation plan artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Risk and dependency register. Objective: Build a practical, auditable deliverable for risk and dependency register in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete risk and dependency register artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Deliverable definition. Objective: Build a practical, auditable deliverable for deliverable definition in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete deliverable definition artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Acceptance criteria author. Objective: Build a practical, auditable deliverable for acceptance criteria author in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete acceptance criteria author artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Guideline first draft. Objective: Build a practical, auditable deliverable for guideline first draft in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete guideline first draft artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Positive and negative examples. Objective: Build a practical, auditable deliverable for positive and negative examples in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete positive and negative examples artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Decision tree writer. Objective: Build a practical, auditable deliverable for decision tree writer in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete decision tree writer artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Boundary cases catalog. Objective: Build a practical, auditable deliverable for boundary cases catalog in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete boundary cases catalog artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Ambiguous case policy. Objective: Build a practical, auditable deliverable for ambiguous case policy in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete ambiguous case policy artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Unknown and abstain policy. Objective: Build a practical, auditable deliverable for unknown and abstain policy in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete unknown and abstain policy artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Multi-label rules. Objective: Build a practical, auditable deliverable for multi-label rules in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete multi-label rules artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Span boundary guide. Objective: Build a practical, auditable deliverable for span boundary guide in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete span boundary guide artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Guideline version changelog. Objective: Build a practical, auditable deliverable for guideline version changelog in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete guideline version changelog artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Annotator FAQ author. Objective: Build a practical, auditable deliverable for annotator faq author in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete annotator faq author artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Sentiment labeling. Objective: Build a practical, auditable deliverable for sentiment labeling in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete sentiment labeling artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Intent tagging. Objective: Build a practical, auditable deliverable for intent tagging in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete intent tagging artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Named entity recognition. Objective: Define entity classes, annotate exact character spans, and explain overlap and boundary rules.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete named entity recognition artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Topic classification. Objective: Build a practical, auditable deliverable for topic classification in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete topic classification artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Toxicity assessment. Objective: Build a practical, auditable deliverable for toxicity assessment in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete toxicity assessment artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: PII identification. Objective: Recognize personally identifying data with conservative redaction; avoid exposing raw sensitive samples.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete pii identification artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Claim and evidence tagging. Objective: Build a practical, auditable deliverable for claim and evidence tagging in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete claim and evidence tagging artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Translation quality tagging. Objective: Build a practical, auditable deliverable for translation quality tagging in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete translation quality tagging artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Duplicate text detection. Objective: Build a practical, auditable deliverable for duplicate text detection in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete duplicate text detection artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Conversation turn labeling. Objective: Build a practical, auditable deliverable for conversation turn labeling in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete conversation turn labeling artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Image classification rubric. Objective: Build a practical, auditable deliverable for image classification rubric in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete image classification rubric artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Bounding box instructions. Objective: Build a practical, auditable deliverable for bounding box instructions in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete bounding box instructions artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Segmentation QA criteria. Objective: Build a practical, auditable deliverable for segmentation qa criteria in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete segmentation qa criteria artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: OCR transcription guide. Objective: Build a practical, auditable deliverable for ocr transcription guide in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete ocr transcription guide artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Video event timestamping. Objective: Build a practical, auditable deliverable for video event timestamping in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete video event timestamping artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Audio transcription QA. Objective: Build a practical, auditable deliverable for audio transcription qa in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete audio transcription qa artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Speaker diarization protocol. Objective: Build a practical, auditable deliverable for speaker diarization protocol in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete speaker diarization protocol artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Visual ambiguity resolver. Objective: Build a practical, auditable deliverable for visual ambiguity resolver in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete visual ambiguity resolver artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Multimodal consistency check. Objective: Build a practical, auditable deliverable for multimodal consistency check in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete multimodal consistency check artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Accessibility description evaluator. Objective: Build a practical, auditable deliverable for accessibility description evaluator in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete accessibility description evaluator artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Instruction following rubric. Objective: Build a practical, auditable deliverable for instruction following rubric in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete instruction following rubric artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Helpfulness evaluation. Objective: Build a practical, auditable deliverable for helpfulness evaluation in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete helpfulness evaluation artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Factuality and citation check. Objective: Build a practical, auditable deliverable for factuality and citation check in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete factuality and citation check artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Relevance grading. Objective: Build a practical, auditable deliverable for relevance grading in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete relevance grading artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Completeness grading. Objective: Build a practical, auditable deliverable for completeness grading in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete completeness grading artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Clarity and style grading. Objective: Build a practical, auditable deliverable for clarity and style grading in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete clarity and style grading artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Pairwise preference rating. Objective: Compare response A and B using grounded criteria without favoring length or position.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete pairwise preference rating artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Refusal appropriateness rubric. Objective: Build a practical, auditable deliverable for refusal appropriateness rubric in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete refusal appropriateness rubric artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Tool-call assessment. Objective: Build a practical, auditable deliverable for tool-call assessment in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete tool-call assessment artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Structured output validity. Objective: Build a practical, auditable deliverable for structured output validity in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete structured output validity artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Preference pair construction. Objective: Build a practical, auditable deliverable for preference pair construction in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete preference pair construction artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Preference rationale writer. Objective: Build a practical, auditable deliverable for preference rationale writer in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete preference rationale writer artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Tie-breaking policy. Objective: Build a practical, auditable deliverable for tie-breaking policy in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete tie-breaking policy artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Reward signal taxonomy. Objective: Build a practical, auditable deliverable for reward signal taxonomy in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete reward signal taxonomy artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Conversation ranking. Objective: Build a practical, auditable deliverable for conversation ranking in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete conversation ranking artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Annotator bias audit. Objective: Build a practical, auditable deliverable for annotator bias audit in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete annotator bias audit artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Length bias diagnostic. Objective: Build a practical, auditable deliverable for length bias diagnostic in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete length bias diagnostic artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Position bias diagnostic. Objective: Build a practical, auditable deliverable for position bias diagnostic in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete position bias diagnostic artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Preference dataset QA. Objective: Build a practical, auditable deliverable for preference dataset qa in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete preference dataset qa artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Feedback synthesis. Objective: Build a practical, auditable deliverable for feedback synthesis in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete feedback synthesis artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Schema validator. Objective: Build a practical, auditable deliverable for schema validator in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete schema validator artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Missing data audit. Objective: Build a practical, auditable deliverable for missing data audit in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete missing data audit artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Duplicate record audit. Objective: Build a practical, auditable deliverable for duplicate record audit in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete duplicate record audit artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Class imbalance analysis. Objective: Build a practical, auditable deliverable for class imbalance analysis in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete class imbalance analysis artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Data leakage screening. Objective: Build a practical, auditable deliverable for data leakage screening in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete data leakage screening artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Train test split review. Objective: Build a practical, auditable deliverable for train test split review in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete train test split review artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Normalization plan. Objective: Build a practical, auditable deliverable for normalization plan in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete normalization plan artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Outlier triage. Objective: Build a practical, auditable deliverable for outlier triage in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete outlier triage artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Metadata consistency check. Objective: Build a practical, auditable deliverable for metadata consistency check in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete metadata consistency check artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Dataset card writer. Objective: Build a practical, auditable deliverable for dataset card writer in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete dataset card writer artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Gold item generation plan. Objective: Build a practical, auditable deliverable for gold item generation plan in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete gold item generation plan artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Gold answer adjudication. Objective: Build a practical, auditable deliverable for gold answer adjudication in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete gold answer adjudication artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Inter-annotator agreement plan. Objective: Choose Cohen kappa, Fleiss kappa, or Krippendorff alpha based on number of raters and label type; note limits.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete inter-annotator agreement plan artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Confusion matrix analysis. Objective: Use supplied ground truth and predictions to compute per-class errors; never invent counts.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete confusion matrix analysis artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Disagreement clustering. Objective: Build a practical, auditable deliverable for disagreement clustering in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete disagreement clustering artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Calibration exercise. Objective: Build a practical, auditable deliverable for calibration exercise in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete calibration exercise artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Blind review protocol. Objective: Build a practical, auditable deliverable for blind review protocol in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete blind review protocol artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Sampling-based QA. Objective: Build a practical, auditable deliverable for sampling-based qa in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete sampling-based qa artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Error severity taxonomy. Objective: Build a practical, auditable deliverable for error severity taxonomy in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete error severity taxonomy artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Quality improvement plan. Objective: Build a practical, auditable deliverable for quality improvement plan in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete quality improvement plan artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Sensitive data inventory. Objective: Build a practical, auditable deliverable for sensitive data inventory in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete sensitive data inventory artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: De-identification protocol. Objective: Build a practical, auditable deliverable for de-identification protocol in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete de-identification protocol artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Consent and licensing checklist. Objective: Build a practical, auditable deliverable for consent and licensing checklist in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete consent and licensing checklist artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Harmful content handling guide. Objective: Build a practical, auditable deliverable for harmful content handling guide in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete harmful content handling guide artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Escalation trigger author. Objective: Build a practical, auditable deliverable for escalation trigger author in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete escalation trigger author artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Access control checklist. Objective: Build a practical, auditable deliverable for access control checklist in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete access control checklist artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Data retention review. Objective: Build a practical, auditable deliverable for data retention review in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete data retention review artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Cultural fairness evaluation. Objective: Build a practical, auditable deliverable for cultural fairness evaluation in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete cultural fairness evaluation artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Adversarial sample handling. Objective: Build a practical, auditable deliverable for adversarial sample handling in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete adversarial sample handling artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Incident report template. Objective: Build a practical, auditable deliverable for incident report template in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete incident report template artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Spreadsheet labeling schema. Objective: Build a practical, auditable deliverable for spreadsheet labeling schema in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete spreadsheet labeling schema artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Label Studio project plan. Objective: Propose labeling configuration, data format, user roles, and QA checks; do not claim configuration has run.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete label studio project plan artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: CVAT task setup checklist. Objective: Specify classes, shapes, interpolation, reviewer roles, and export verification.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete cvat task setup checklist artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Annotation queue design. Objective: Build a practical, auditable deliverable for annotation queue design in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete annotation queue design artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Batch assignment planner. Objective: Build a practical, auditable deliverable for batch assignment planner in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete batch assignment planner artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Automation opportunity audit. Objective: Build a practical, auditable deliverable for automation opportunity audit in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete automation opportunity audit artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Webhook payload schema. Objective: Build a practical, auditable deliverable for webhook payload schema in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete webhook payload schema artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Export validation plan. Objective: Build a practical, auditable deliverable for export validation plan in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete export validation plan artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Progress dashboard specification. Objective: Build a practical, auditable deliverable for progress dashboard specification in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete progress dashboard specification artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Operational handoff SOP. Objective: Build a practical, auditable deliverable for operational handoff sop in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete operational handoff sop artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Daily production summary. Objective: Build a practical, auditable deliverable for daily production summary in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete daily production summary artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Weekly quality report. Objective: Build a practical, auditable deliverable for weekly quality report in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete weekly quality report artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Error analysis memo. Objective: Build a practical, auditable deliverable for error analysis memo in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete error analysis memo artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Guideline update proposal. Objective: Build a practical, auditable deliverable for guideline update proposal in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete guideline update proposal artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Stakeholder question log. Objective: Build a practical, auditable deliverable for stakeholder question log in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete stakeholder question log artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Metrics interpretation. Objective: Build a practical, auditable deliverable for metrics interpretation in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete metrics interpretation artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Root cause analysis. Objective: Build a practical, auditable deliverable for root cause analysis in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete root cause analysis artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Remediation tracking. Objective: Build a practical, auditable deliverable for remediation tracking in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete remediation tracking artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Final dataset handoff. Objective: Build a practical, auditable deliverable for final dataset handoff in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete final dataset handoff artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Client-ready executive summary. Objective: Build a practical, auditable deliverable for client-ready executive summary in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete client-ready executive summary artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Beginner practice dataset. Objective: Build a practical, auditable deliverable for beginner practice dataset in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete beginner practice dataset artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: NER portfolio case study. Objective: Build a practical, auditable deliverable for ner portfolio case study in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete ner portfolio case study artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Sentiment QA project. Objective: Build a practical, auditable deliverable for sentiment qa project in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete sentiment qa project artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Image labeling showcase. Objective: Build a practical, auditable deliverable for image labeling showcase in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete image labeling showcase artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: LLM evaluation sample. Objective: Build a practical, auditable deliverable for llm evaluation sample in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete llm evaluation sample artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Preference annotation portfolio. Objective: Build a practical, auditable deliverable for preference annotation portfolio in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete preference annotation portfolio artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Quality dashboard project. Objective: Build a practical, auditable deliverable for quality dashboard project in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete quality dashboard project artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Annotation resume bullets. Objective: Build a practical, auditable deliverable for annotation resume bullets in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete annotation resume bullets artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Remote interview simulation. Objective: Build a practical, auditable deliverable for remote interview simulation in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete remote interview simulation artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.

Act as a senior AI data annotation and quality-assurance specialist. Task: Freelance proposal drafter. Objective: Build a practical, auditable deliverable for freelance proposal drafter in a professional annotation project.

PROJECT INPUTS (replace brackets before use):
• Domain/use case: [DOMAIN_AND_USE_CASE]
• Dataset modality and schema: [TEXT_IMAGE_AUDIO_VIDEO_OR_MULTIMODAL_AND_FIELDS]
• Label taxonomy or evaluation criteria: [LABELS_AND_DEFINITIONS]
• Sample records or anonymized examples: [SAMPLES_OR_NONE]
• Intended audience / stakeholders: [AUDIENCE]
• Constraints, locale, policy, and tooling: [CONSTRAINTS]

WORK INSTRUCTIONS:
1. Restate the task, assumptions, and exact unit of annotation. Ask up to 3 blocking questions only if essential; otherwise proceed with clearly marked assumptions.
2. Produce the concrete freelance proposal drafter artifact tailored to the supplied labels, modality, and stakeholders—not generic advice.
3. Define an operational step-by-step process with decision rules, positive/negative cases, edge cases, and an abstain/escalate route when appropriate.
4. Include at least 3 realistic, clearly SYNTHETIC examples with expected labels or outcomes and short adjudication rationales. Do not fabricate evaluation of unseen data.
5. Specify measurable QA checks, acceptance thresholds as PROPOSED (not observed), common failure modes, and corrective actions.
6. Identify potential privacy, copyright, safety, fairness, and data leakage risks relevant to this task; minimize exposure of sensitive data.
7. Deliver a reusable table, checklist, rubric, CSV header, or JSON schema where appropriate, plus handoff notes.

OUTPUT FORMAT: A. Goal and inputs; B. Definitions and method; C. Actionable artifact; D. Synthetic examples; E. Edge cases and escalation; F. QA and metrics; G. Risks and review; H. Ready-to-use checklist.
Do not claim that any tool was run, any dataset was inspected, or any metric was measured unless those results are actually provided. Distinguish recommendations from facts.