Human Handoff Trigger Prompt for Low-Confidence Responses

This prompt is a post-generation formatting tool for support operations engineers and review-queue designers. Its job is to convert a low-confidence model response into a structured JSON handoff payload for a human agent. Use it when your primary model has already generated a draft answer, but the confidence score, uncertainty flags, or safety checks indicate the response is not safe to send autonomously. The prompt does not generate the initial answer, re-evaluate the model's logic, or decide if an escalation should occur. It assumes the escalation decision has already been made by an upstream gate and focuses entirely on packaging the context a human reviewer needs to act efficiently.

Before invoking this prompt, you must have the following inputs available: the original user query, the model's draft response, a confidence score (typically a float between 0.0 and 1.0), a list of specific uncertainty flags or violated safety rules, and a reason code for the escalation. The prompt template uses these as [ORIGINAL_QUERY], [DRAFT_RESPONSE], [CONFIDENCE_SCORE], [UNCERTAINTY_FLAGS], and [ESCALATION_REASON]. The output is a strict JSON payload containing these fields plus a generated escalation_id and timestamp. This structure ensures that human reviewers receive complete context in a single, predictable object without hunting through logs, trace views, or raw model outputs.

Do not use this prompt for generating the initial answer, for re-ranking confidence scores, or for making the binary decision to escalate. Those functions belong to separate prompts in the Confidence and Uncertainty Escalation group, such as the Threshold-Based Escalation Decision Prompt or the Confidence Score Calibration Prompt Template. Also avoid this prompt when the model's draft is empty or the confidence score is above your production threshold for autonomous delivery. In those cases, the handoff payload would contain no actionable content for the reviewer and would waste human cycles. The next section provides the copy-ready template you can adapt to your specific escalation schema and review-queue format.

This prompt template is designed to be injected into your escalation harness immediately after a confidence evaluation step. It takes the original user query, the model's low-confidence draft response, and the confidence breakdown, then produces a structured handoff payload suitable for a human review queue. The output is designed to give a reviewer everything they need to make a decision without hunting through logs.

textCopy
SYSTEM: You are an escalation formatter for an AI support system. Your job is to produce a structured handoff record when the primary model's response confidence falls below the acceptable threshold. Do not answer the user's original query. Do not evaluate the draft response. Only format the escalation payload.

USER QUERY:
[USER_QUERY]

MODEL DRAFT RESPONSE:
[DRAFT_RESPONSE]

CONFIDENCE BREAKDOWN:
[CONFIDENCE_BREAKDOWN]

ESCALATION REASON:
[ESCALATION_REASON]

RETRY BUDGET STATE:
[RETRY_BUDGET_STATE]

OUTPUT SCHEMA:
{
  "handoff_id": "string, unique identifier for this escalation event",
  "timestamp": "string, ISO 8601 timestamp of escalation",
  "original_query": "string, the user's original input verbatim",
  "model_draft": "string, the low-confidence draft response exactly as generated",
  "confidence_summary": {
    "overall_score": "number, the aggregate confidence score (0.0 to 1.0)",
    "threshold": "number, the configured threshold that was not met",
    "low_confidence_spans": [
      {
        "claim": "string, the specific claim or segment with low confidence",
        "score": "number, confidence score for this claim",
        "reason": "string, why confidence is low (e.g., missing_evidence, ambiguity, contradiction)"
      }
    ]
  },
  "escalation_reason": "string, the primary reason for escalation",
  "retry_context": {
    "attempts_used": "number, retry attempts consumed before escalation",
    "budget_remaining": "number, retry attempts still available",
    "previous_errors": ["string, error messages from prior attempts if any"]
  },
  "review_questions": ["string, specific questions the human reviewer should answer"],
  "recommended_actions": ["string, suggested next steps for the reviewer"]
}

CONSTRAINTS:
- Do not modify or paraphrase the original query or model draft.
- Include all fields in the output schema, using null or empty arrays where no data exists.
- Review questions must be specific to the confidence gaps identified, not generic.
- If the retry budget is exhausted, note this explicitly in recommended_actions.

To adapt this template, replace each square-bracket placeholder with values from your confidence evaluation step. The [CONFIDENCE_BREAKDOWN] should include per-claim scores if your evaluation pipeline produces them; otherwise, provide the aggregate score and any available detail. The [ESCALATION_REASON] should come from a predefined enum in your system (e.g., below_threshold, evidence_missing, ambiguity_unresolved, retry_budget_exhausted). Before deploying, validate that the output JSON parses correctly and that all required fields are present—missing review_questions is a common failure mode that leaves reviewers without clear guidance.

This prompt is designed to be the final gate in an AI pipeline before a response reaches a user. It should be invoked when a primary model's output has already been flagged for low confidence, either by an upstream classifier, a structured confidence score below a configured threshold, or a validator that detected missing citations. The prompt's job is not to re-answer the question but to produce a structured, machine-readable handoff payload that a downstream review queue or orchestration layer can consume without parsing free text.

Wire this prompt into your application as a synchronous step within an if confidence < threshold block. The input variables [ORIGINAL_QUERY], [MODEL_DRAFT], [CONFIDENCE_BREAKDOWN], and [ESCALATION_REASON] must be populated by the upstream system before this prompt is called. The output must be validated against a strict JSON schema before the handoff is accepted. If the model fails to produce valid JSON after one retry with a schema error message injected into [PREVIOUS_ERROR], abort the retry loop and escalate to a human operator with the raw model output attached. Log every handoff payload, including the handoff_id, timestamp, and escalation_path, to an append-only audit table for later review.

For high-risk domains where a missed escalation could cause harm, do not rely solely on the model's self-assessment. Implement a secondary validation step: check that the completeness_checks array in the output is non-empty and that each check has a boolean passed value. If any completeness check fails, force the handoff regardless of the model's escalation_decision. This prompt is a safety net, not a decision-maker. Its primary value is in structuring the handoff context so that a human reviewer can quickly understand why the AI stopped and what it was trying to do.