Decision-Ready Brief Prompt for Human Reviewers

This prompt is designed for operations leads and platform engineers who need to convert raw AI outputs, flagged items, or agent-generated work into a concise brief that a human reviewer can act on immediately. The job-to-be-done is not to dump data into a queue; it is to present pre-computed options, trade-offs, and a recommended action so the reviewer spends seconds understanding the situation and seconds deciding, not minutes reconstructing context. The ideal user is someone building a human-in-the-loop system where the AI has already gathered evidence, identified a decision point, and needs a human to choose among bounded options. The required context includes the original task, the evidence collected, the options considered, and the AI's confidence in each option.

Use this prompt when your system has already completed the analysis phase and the remaining work is a constrained choice. For example, a content moderation system that has flagged a post, identified the specific policy it might violate, and needs a human to confirm the violation and choose between 'remove', 'warn', or 'ignore' is a perfect fit. Another strong use case is a code review agent that has found a potential security vulnerability, linked to the specific line of code and the CWE, and needs a human to approve the fix or mark it as a false positive. The prompt works because it forces the AI to structure its output around the decision, not the discovery.

Do not use this prompt when the situation is entirely novel, when the options are unbounded, or when the AI should act autonomously without human input. If the AI cannot enumerate the likely next steps with reasonable confidence, this prompt will produce a misleadingly tidy brief that hides critical uncertainty. For open-ended investigations, use a 'Context Summary Prompt for Human Reviewer Handoff' instead. If the AI is authorized to act without human approval, this prompt adds unnecessary latency. The key test is whether a human reviewer, seeing only the brief, can make a confident decision without needing to re-verify the underlying evidence from scratch. If they can't, the prompt is being used in the wrong place, and you should fall back to a more detailed evidence-packaging prompt.

The prompt below is designed to be wired directly into your review queue pipeline. It takes a work item, any relevant context, and your output schema, then instructs the model to produce a decision-ready brief. The brief is not a narrative summary; it is a structured artifact containing pre-computed options, explicit trade-offs, and a single recommended action. This format lets human reviewers scan, evaluate, and decide quickly without reconstructing the problem from scratch.

textCopy
You are a decision-support assistant preparing briefs for human reviewers. Your output must be structured, scannable, and actionable. Do not write narrative summaries. Present options, trade-offs, and a clear recommendation.

## INPUT
Work Item: [WORK_ITEM]
Context: [CONTEXT]

## OUTPUT SCHEMA
Return a JSON object with the following structure:
{
  "brief_id": "string",
  "summary": "One-sentence summary of the decision required.",
  "options": [
    {
      "id": "string",
      "label": "Short option name",
      "description": "What this option entails.",
      "pros": ["string"],
      "cons": ["string"],
      "risk_level": "low | medium | high | critical",
      "preconditions": ["string"]
    }
  ],
  "recommendation": {
    "option_id": "string",
    "rationale": "Why this option is recommended over others.",
    "confidence": "high | medium | low",
    "key_uncertainties": ["string"]
  },
  "information_gaps": ["What is missing that would increase confidence."],
  "escalation_triggers": ["Conditions that should cause immediate escalation."]
}

## CONSTRAINTS
- Generate at least 2 and no more than 5 options.
- Every option must include at least one pro and one con.
- The recommendation must reference a specific option_id.
- If confidence is "low," the brief must include at least one information_gap.
- Do not invent facts not present in the input context.
- If the work item involves [RISK_LEVEL] risk, flag it in escalation_triggers.

## EXAMPLES
[EXAMPLES]

## TOOLS
[TOOLS]

To adapt this template, replace the square-bracket placeholders with your application's runtime values. [WORK_ITEM] should contain the full text of the item requiring review. [CONTEXT] should include relevant evidence, policy documents, or historical decisions. [RISK_LEVEL] is a string you set based on your internal risk taxonomy. [EXAMPLES] should be replaced with one or two few-shot demonstrations of correct briefs for your domain. [TOOLS] can be omitted or replaced with available function definitions if the model needs to fetch additional data. After generation, validate the JSON structure before presenting it to a reviewer. If the model omits required fields or produces a confidence level inconsistent with the information gaps, reject the output and retry with a more explicit constraint in the prompt.

The Decision-Ready Brief prompt is designed to sit at the boundary between an AI agent's autonomous execution and a human reviewer's queue. It should be invoked after the agent has gathered evidence, identified possible actions, and determined that a decision requires human judgment. The prompt's output is a structured brief that becomes the primary payload for a review item in your queue system. Do not use this prompt for actions the agent can safely resolve alone; it is specifically for pre-decision synthesis when escalation is required.

To wire this into an application, wrap the prompt call in a function that accepts the agent's current state, including the original objective, gathered evidence, and any tool outputs. The function should inject these into the [CONTEXT] and [EVIDENCE] placeholders. After receiving the model's response, validate the output against a strict JSON schema that enforces the presence of decision_options, trade_offs, recommended_action, and missing_information fields. If validation fails, implement a single retry with the validation error injected into the [CONSTRAINTS] field. After a second failure, log the raw output and flag the item for manual triage rather than entering an infinite retry loop. For high-risk domains, always set [RISK_LEVEL] to high, which will instruct the model to include blocking conditions and irreversible consequence flags in the brief.

Model choice matters here. Use a model with strong instruction-following and structured output capabilities, such as gpt-4o or claude-3.5-sonnet. Avoid smaller or older models that may drop fields or hallucinate options. For auditability, log the full prompt, the model's raw response, the validated brief, and the reviewer's eventual decision in an immutable audit trail. This creates a traceable record from agent action to human override. Do not allow the brief to be surfaced to the reviewer without validation; an unvalidated brief can omit critical trade-offs or frame options in a biased way, leading to poor decisions. Finally, ensure your review queue UI renders the decision_options as distinct, selectable buttons and displays missing_information prominently so reviewers can quickly identify gaps before committing to a decision.