Human Review Handoff Packet Generation Prompt

This prompt is designed for verification pipeline operators and triage system architects who need to produce a structured, review-ready packet when an automated fact-checking system cannot resolve a claim with sufficient confidence. The core job-to-be-done is bridging the gap between machine-scale verification and human judgment: the system has already extracted claims, retrieved evidence, and attempted automated matching, but the result is ambiguous, contradictory, or falls below a confidence threshold. Instead of silently discarding the claim or returning a misleading 'unverified' label, this prompt generates a complete handoff packet that gives a human reviewer everything they need to make a fast, informed decision without redoing the machine's work.

Use this prompt when your verification pipeline has already completed claim extraction, evidence retrieval, and preliminary scoring, and the claim meets at least one of these escalation criteria: confidence score below threshold, contradictory evidence from multiple sources, high-stakes domain (legal, medical, financial, safety), novel or edge-case claim type the model was not calibrated for, or explicit regulatory or compliance requirement for human review. The prompt expects structured inputs including the original claim, retrieved evidence with source metadata, the AI's preliminary assessment and confidence score, and any relevant context such as the source document or user query that triggered the verification. Do not use this prompt for claims that can be auto-verified with high confidence, for real-time chat moderation where latency prohibits human review, or as a substitute for building proper confidence thresholds and triage logic in your application layer.

This prompt is not a general-purpose summarizer, a chatbot response generator, or a replacement for your verification pipeline's core logic. It assumes you already have a working claim extraction and evidence retrieval system. If you are still building that pipeline, start with the Claim Extraction and Decomposition prompts and Evidence Matching and Source Attribution prompts in this pillar before implementing human review handoffs. The output of this prompt is a structured packet intended for ingestion into a review queue, ticketing system, or compliance audit trail—not a conversational reply. Before deploying, ensure your application validates the output schema, logs every handoff for auditability, and routes packets to reviewers with appropriate domain expertise for the claim type.

This prompt template produces a structured handoff packet that equips a human reviewer with everything they need to assess an ambiguous claim: the original claim in context, available evidence, the AI's preliminary assessment, and specific, actionable questions. The template is designed to be dropped into a triage system where automated verification has reached its confidence boundary. Use square-bracket placeholders to inject claim data, evidence records, and routing metadata at runtime. The output is a self-contained review packet, not a final verdict.

textCopy
You are a verification triage assistant. Your job is to prepare a structured handoff packet for a human reviewer when an automated fact-checking pipeline cannot reach a high-confidence verdict on a claim.

Generate a handoff packet using the exact structure below. Do not include a final verdict. Your output must help the reviewer reach a decision efficiently.

## Input
- Claim: [CLAIM_TEXT]
- Claim ID: [CLAIM_ID]
- Source Context: [SOURCE_CONTEXT]
- Retrieved Evidence (JSON array of evidence objects with fields: source_id, source_name, source_authority_score, excerpt, relevance_to_claim, supports_claim [true/false/null]): [EVIDENCE_ARRAY]
- AI Preliminary Assessment: [AI_PRELIM_ASSESSMENT]
- Confidence Score (0.0-1.0): [CONFIDENCE_SCORE]
- Reason for Escalation: [ESCALATION_REASON]
- Reviewer Role: [REVIEWER_ROLE]
- Deadline (ISO 8601 or null): [DEADLINE]

## Output Schema
Produce a JSON object with these fields:
{
  "packet_id": "string, unique identifier for this handoff",
  "generated_at": "ISO 8601 timestamp",
  "claim": {
    "text": "string, the claim under review",
    "id": "string",
    "source_context": "string, surrounding text or metadata where the claim appeared"
  },
  "escalation": {
    "reason": "string, why automated verification could not resolve this claim",
    "confidence_score": "number, the AI's confidence before escalation",
    "preliminary_assessment": "string, the AI's tentative finding and its limitations"
  },
  "evidence_packet": {
    "total_sources": "number",
    "supporting_sources": "number",
    "contradicting_sources": "number",
    "inconclusive_sources": "number",
    "sources": [
      {
        "source_id": "string",
        "source_name": "string",
        "authority_score": "number",
        "excerpt": "string, relevant passage",
        "relevance": "string, why this source matters",
        "direction": "supports | contradicts | inconclusive"
      }
    ]
  },
  "reviewer_questions": [
    "string, specific, actionable question the reviewer should answer"
  ],
  "reviewer_guidance": {
    "role": "string, intended reviewer role",
    "deadline": "ISO 8601 or null",
    "priority": "high | medium | low",
    "suggested_actions": ["string, concrete next step for the reviewer"]
  },
  "context_completeness_check": {
    "missing_context_types": ["string, types of context not available"],
    "evidence_gaps": ["string, specific evidence that would help but is absent"],
    "completeness_notes": "string, overall assessment of packet completeness"
  },
  "bias_minimization_notes": {
    "potential_biases": ["string, identified risks of bias in evidence or assessment"],
    "mitigation_suggestions": ["string, how the reviewer can counter these biases"]
  }
}

## Constraints
- Do not fabricate evidence. Only use evidence provided in [EVIDENCE_ARRAY].
- If evidence is missing or insufficient, state that explicitly in context_completeness_check and evidence_gaps.
- Reviewer questions must be specific and answerable. Avoid vague questions like "Is this true?"
- The preliminary_assessment field must include the AI's uncertainty and limitations, not just a tentative verdict.
- If the claim touches regulated domains (legal, medical, financial), add a note in reviewer_guidance.suggested_actions recommending consultation with a domain expert.
- Set priority based on [CONFIDENCE_SCORE] and [ESCALATION_REASON]: scores below 0.3 with high-impact claims are high priority; scores above 0.7 with minor claims are low priority.
- Do not include a final verdict field. This packet is for review, not resolution.

To adapt this template, replace each square-bracket placeholder with runtime data from your verification pipeline. The [EVIDENCE_ARRAY] should be a serialized JSON array matching the evidence object structure described in the prompt. If your system uses different evidence fields, update both the input description and the output schema's evidence_packet.sources structure to match. The [AI_PRELIM_ASSESSMENT] should be the raw output from your upstream verification step, including its confidence score and reasoning. Before deploying, validate that the generated JSON matches the schema and that reviewer_questions are non-empty and specific. For high-stakes domains, add a post-generation validation step that checks for missing bias_minimization_notes and context_completeness_check fields before the packet reaches a human reviewer.

This prompt sits at the boundary between automated verification and human judgment. It should be called only after upstream classifiers have determined that a claim cannot be auto-resolved—typically because confidence scores fall into a 'review required' band, evidence is contradictory, or the claim domain is high-risk. The prompt expects a structured input payload containing the claim, available evidence, AI preliminary assessment, and routing metadata. Do not call this prompt for claims that can be auto-verified or auto-rejected; doing so wastes reviewer time and creates noise in the review queue.

Wire the prompt into a queue-based workflow. When a claim is flagged for human review, assemble the input payload from upstream pipeline outputs: the original claim text, the evidence set with source identifiers and relevance scores, the AI's preliminary verdict and confidence, and any contradiction flags. Pass this payload into the prompt with a strict output schema expectation. Validate the returned JSON immediately: check that all required fields are present, that the reviewer_questions array is non-empty, that source references in the packet match the provided evidence set, and that the context_completeness and bias_minimization fields contain substantive content rather than placeholder text. If validation fails, retry once with the validation errors appended to the prompt as feedback. If the second attempt also fails, log the failure and escalate to a system administrator rather than routing a malformed packet to a reviewer.

Model choice matters here. Use a model with strong instruction-following and structured output capabilities. The prompt includes bias-minimization instructions that require the model to reflect on its own preliminary assessment; weaker models may produce superficial self-critique or generic disclaimers. For high-stakes domains such as healthcare or legal review, route the generated packet through a second validation step that checks for leading questions, unbalanced evidence presentation, or premature conclusions in the AI preliminary assessment section. Log every generated packet with a unique handoff ID, timestamp, model version, and validation status before it reaches the reviewer queue. This audit trail is essential for governance and for diagnosing systematic handoff quality issues over time.

Avoid wiring this prompt directly into a real-time user-facing interface without a review buffer. Human reviewers need time to process packets, and generating packets faster than reviewers can handle them creates a backlog that undermines the triage system's purpose. Implement a queue depth monitor that throttles packet generation when the review queue exceeds capacity. Also avoid using this prompt for claims where the evidence set is empty or trivially insufficient; in those cases, route directly to an 'insufficient evidence' disclosure workflow rather than consuming reviewer time with a packet that contains no actionable evidence to review.