Structured Uncertainty Packaging Prompt for Human Review

This prompt is designed for operations teams and AI reliability engineers who need to route ambiguous or low-confidence AI outputs to a human review queue. Instead of letting the model guess or fail silently, this prompt packages the original input, the model's output, a confidence breakdown, identified evidence gaps, and specific questions for the human reviewer into a single structured payload. The core job-to-be-done is creating an efficient, auditable handoff from an autonomous system to a human operator when the system recognizes it is operating beyond its reliable capability. The ideal user is someone building a production AI pipeline where the cost of an incorrect autonomous action—such as a wrong classification, a hallucinated fact in a customer response, or an invalid API call—is high enough to warrant human intervention.

Use this prompt when your system has already determined that a model's output is unreliable, typically because a confidence score has fallen below a defined threshold. It is the step after uncertainty detection, not the detection mechanism itself. You should integrate it into workflows where the model's output requires human judgment before it can be acted upon, such as reviewing a flagged support ticket, verifying a generated code change, or approving a financial transaction summary. The prompt is not for generating user-facing answers; it is an internal escalation mechanism. It is also not a replacement for a confidence scoring system—you must have a confidence signal (from logprobs, a separate classifier, or a self-assessment prompt) to trigger this packaging step. Avoid using this prompt for low-stakes decisions where the cost of review exceeds the cost of an error, or for real-time user interactions where a human cannot be looped in within an acceptable latency window.

Before implementing, ensure you have a defined review queue and a clear SLA for human reviewers. The prompt's output is only as valuable as the reviewer's ability to act on it quickly. A common failure mode is packaging too much information, which overwhelms the reviewer and slows down the queue. To prevent this, constrain the [EVIDENCE_GAPS] and [REVIEWER_QUESTIONS] sections to only the most critical items. The next step after reading this section is to examine the prompt template and adapt its output schema to match your review queue's exact data contract. If your primary need is to detect uncertainty rather than package it for review, start with the 'Low-Confidence Classification Escalation Prompt Template' or 'Model Confidence Score Extraction Prompt' instead.

This prompt template transforms a low-confidence or ambiguous AI output into a complete review package for a human operator. Instead of forcing the model to guess or returning a vague disclaimer, the prompt instructs the model to compile the original input, its own output, a confidence breakdown, identified evidence gaps, and specific, actionable questions for the reviewer. The goal is to minimize the time a human spends reconstructing context before they can make a decision.

textCopy
SYSTEM: You are an AI system packaging uncertain outputs for human review. Your job is not to decide, but to prepare a complete, structured brief that lets a human reviewer decide quickly and accurately.

USER:
Package the following AI output for human review. The output was flagged because [ESCALATION_REASON].

ORIGINAL INPUT:
[USER_INPUT]

MODEL OUTPUT:
[MODEL_OUTPUT]

CONFIDENCE ASSESSMENT:
[CONFIDENCE_DETAILS]

INSTRUCTIONS:
1. Summarize the original request and the model's response in 1-2 sentences.
2. Break down the confidence assessment into specific claims, each with a confidence level (High/Medium/Low) and a brief reason for the rating.
3. List the specific evidence or information gaps that prevent high confidence.
4. Formulate 2-4 specific, non-leading questions for the human reviewer. Each question should target a single decision point.
5. If the output involves a proposed action, state the action, its reversibility, and the risk of proceeding vs. stopping.
6. Format the entire package using the following structure:

REVIEW PACKAGE
---
SUMMARY: [1-2 sentence summary]

CONFIDENCE BREAKDOWN:
- Claim: [claim text]
  Confidence: [High/Medium/Low]
  Reason: [reason for rating]
- ...

EVIDENCE GAPS:
- [gap description]
- ...

REVIEWER QUESTIONS:
1. [specific question]
2. ...

PROPOSED ACTION (if any):
- Action: [action description]
- Reversibility: [Reversible/Irreversible/Conditionally Reversible]
- Risk if Proceeded: [risk description]
- Risk if Stopped: [risk description]

OUTPUT CONSTRAINTS:
- Do not add new information or analysis beyond what was in the original model output.
- Do not answer the reviewer questions yourself.
- If the confidence assessment is missing or malformed, flag it as an incomplete package and request a re-run.

To adapt this template, replace the square-bracket placeholders with data from your application's orchestration layer. [ESCALATION_REASON] should be a machine-readable tag (e.g., CONFIDENCE_BELOW_THRESHOLD, CONFLICTING_EVIDENCE, OOD_INPUT) that your routing logic can use. [CONFIDENCE_DETAILS] should be the raw output of your confidence scoring step, passed in as structured text. The output structure is designed to be parsed by a downstream review queue UI, so keep the REVIEW PACKAGE header and field labels stable. Before deploying, validate that the generated package is parseable and that reviewer questions are genuinely decision-oriented, not rhetorical.

This prompt is designed to sit at the boundary between an AI system and a human review queue. It should be invoked when a primary model's output fails a confidence threshold, an ambiguity detector flags the input, or a downstream validator rejects the result. The prompt's job is not to decide the outcome but to package the unresolved state into a review-ready payload that minimizes the time a human reviewer spends reconstructing context. In a production harness, this prompt is typically called by an orchestrator or a guardrail function, not directly by an end user.

The implementation should enforce a strict contract around the prompt's inputs and outputs. Before calling the prompt, the harness must assemble the [ORIGINAL_INPUT], [MODEL_OUTPUT], [CONFIDENCE_BREAKDOWN], [EVIDENCE_GAPS], and [ESCALATION_REASON] from the upstream system's state. After receiving the response, validate the output against a JSON schema that requires the review_package object with summary, specific_questions, and recommended_reviewer_role fields. If validation fails, retry once with a repair prompt that includes the schema violation errors. If the retry also fails, log the raw output and escalate to an on-call channel with a review_package_generation_failed event. For model choice, a mid-tier model like claude-3-5-haiku or gpt-4o-mini is usually sufficient since the task is structured packaging, not complex reasoning. Reserve more powerful models for cases where the evidence gaps require nuanced summarization.

Logging and observability are critical because this prompt is the handoff point between automation and human judgment. Every invocation should emit a structured log with the escalation reason, the confidence scores that triggered the handoff, the generated review package, and the latency of the packaging step. If the review package is routed to a queue like Jira, Linear, or a custom review tool, include a correlation ID that links back to the original AI transaction. Avoid building this as a synchronous blocking call in user-facing paths; the packaging step should be asynchronous where possible, with the review package delivered to the queue and the user receiving an acknowledgment that a human will review the item. Never allow the system to proceed autonomously after this prompt fires—the whole point is that the AI has reached its safe operating limit.