Handoff Prompt with Confidence Score and Alternatives

Use this prompt when an agent generates multiple candidate answers, classifications, extractions, or action plans and must hand off to a human for final selection. This is common in high-stakes workflows where the agent can narrow the field but should not unilaterally decide: compliance reviews, medical coding, contract clause selection, fraud disposition, or content moderation. The prompt structures the handoff so the human sees ranked options, understands why each is plausible, and can choose with full context. The core job-to-be-done is transforming raw agent output into a decision-ready brief that reduces cognitive load on the reviewer while preserving the evidence they need to exercise judgment.

Do not use this prompt when the agent has a single high-confidence output with no meaningful alternatives, when the decision is low-risk and can be automated, or when the human needs raw agent traces rather than a structured decision brief. This prompt is also inappropriate when the alternatives are trivially similar or when the confidence differential between the top candidate and the next-best is so large that presenting alternatives adds noise rather than value. In those cases, a simpler handoff with a single recommendation and a confidence threshold gate is more appropriate. The prompt assumes the agent has already produced multiple candidates with associated scores; it does not generate the candidates themselves.

Before deploying this prompt, ensure your upstream agent produces at least two distinct outputs with per-candidate confidence scores and a brief rationale for each. The human reviewer should receive enough context to choose without replaying the agent's entire reasoning trace. Wire the output into a review queue or approval UI, and log which candidate was selected for downstream evaluation and model improvement. Avoid using this prompt as a substitute for improving agent accuracy when the root cause is poor retrieval, insufficient context, or a model that is not fit for the task.

This prompt is designed to be injected into your agent's handoff instruction block after it has already generated multiple candidate outputs and assigned a confidence score to each. It structures the final message to a human reviewer so they can make an informed selection without reading raw agent logs. The template assumes the upstream agent has access to a list of candidates, their scores, and the key differentiators between them.

textCopy
You are handing off a decision to a human reviewer. Your goal is to present the top candidates clearly so the reviewer can select the best option with minimal effort.

## Candidates
Here are the candidates you generated, ranked by confidence score:
[CANDIDATE_LIST]

## Confidence Scores
Confidence scores range from 0.0 (no confidence) to 1.0 (certain). Scores are based on [CONFIDENCE_CRITERIA].
[CONFIDENCE_BREAKDOWN]

## Key Differentiators
These are the most important differences between the top candidates:
[DIFFERENTIATORS]

## Recommendation
Based on the evidence and confidence scores, the recommended candidate is [RECOMMENDED_CANDIDATE_ID] because [RECOMMENDATION_RATIONALE].

## Handoff Instructions
Present the following to the human reviewer in this exact structure:

### Summary
A one-sentence summary of the task and what decision is needed.

### Top Candidates
For each candidate in the top [TOP_N] (do not exceed [MAX_CANDIDATES]):
- **Candidate ID**: [ID]
- **Output**: [CANDIDATE_OUTPUT]
- **Confidence**: [SCORE] — [BRIEF_CONFIDENCE_EXPLANATION]
- **Key Strengths**: [STRENGTHS]
- **Key Risks**: [RISKS]

### Recommendation
State the recommended candidate, the primary reason, and any caveats the reviewer should consider.

### Decision Request
Ask the reviewer: "Which candidate should we proceed with?" Include a note that they can request more details on any candidate.

## Constraints
- Do not present more than [MAX_CANDIDATES] candidates.
- If the top candidate's confidence is below [ESCALATION_THRESHOLD], add a warning: "All candidates have low confidence. Consider requesting additional information or manual resolution."
- If two candidates have confidence scores within [TIE_DELTA] of each other, flag them as a near-tie and explain the trade-off.
- Never fabricate confidence scores. If a score is unavailable, mark it as "Not scored" and explain why.
- If [RISK_LEVEL] is "high", append: "This decision has high-risk implications. Reviewer must confirm understanding of consequences before selecting."

To adapt this template, replace each square-bracket placeholder with values from your agent's upstream context. [CANDIDATE_LIST] should be populated from the agent's generation step, [CONFIDENCE_CRITERIA] should describe how scores were computed (e.g., source agreement, retrieval score, model self-assessment), and [ESCALATION_THRESHOLD] should match your system's risk tolerance. For high-stakes domains, set [RISK_LEVEL] to "high" and ensure the human reviewer's selection is logged with a timestamp and reviewer identity for audit trails. Before deploying, test this prompt with at least three candidate sets where the top score is above threshold, below threshold, and in a near-tie to verify the branching logic works correctly.

This prompt is designed to sit at the boundary between an autonomous agent and a human review queue. The agent calls it after generating multiple candidate outputs with varying confidence scores. The harness must validate the structured output, enforce the confidence schema, and route the payload to the correct review interface. Do not treat this as a one-shot generation step; it is a contract between the agent's internal reasoning and the human's decision surface.

Validation and retry logic. Parse the model output against a strict JSON schema that requires candidates (array, min 1, max 5), each with id, summary, confidence (0.0–1.0 float), key_differentiators (array of strings), and risks (array of strings). Also require a top-level recommendation object with selected_id, rationale, and uncertainty_notes. If parsing fails, retry once with the raw output and a repair instruction appended to the prompt. If the second attempt fails, log the failure and route to a human with the raw output attached as unstructured context. If confidence scores are missing or out of range, flag the payload for review with a validation_warning field injected into the handoff record.

Routing and human review integration. After validation, check the top candidate's confidence score against a configurable threshold (default 0.85). If confidence is below threshold, route to a human review queue with the full candidate list and recommendation. If confidence is above threshold, you may auto-accept the recommendation, but always log the decision with the full payload, timestamp, agent ID, and session context for auditability. For high-risk domains, require human approval regardless of confidence. The harness should expose a webhook or API endpoint that the review interface can call to return the human's selection, which the agent then uses as the final output for downstream steps.

Model choice and latency considerations. This prompt benefits from models with strong instruction-following and structured output capabilities. Use a model that supports JSON mode or function calling to reduce parsing failures. If latency is critical, consider a smaller, fine-tuned model for the handoff formatting step while keeping the candidate generation on a more capable model. Cache the system prompt prefix to reduce per-request token costs. Monitor the rate of validation failures and human escalations as key observability signals—rising rates indicate either model drift, input distribution shift, or threshold misconfiguration.