Human-in-the-Loop Triage Prompt Template

This prompt template is designed for support and operations teams that need a reliable, structured way to route tasks from an AI agent to a human reviewer. The core job is triage: transforming the raw, often verbose log of an agent's actions into a concise, prioritized summary that a human can act on immediately. The ideal user is an engineering lead or product manager building a human-in-the-loop system where the cost of a wrong autonomous action is high, such as processing a refund, approving a sensitive account change, or resolving a complex technical support ticket. It is not a general-purpose summarization tool or a replacement for a full case management system.

Use this prompt when an agent has completed its analysis or hit a defined boundary and must hand off a task. The prompt requires several concrete inputs to be effective: a structured [AGENT_ACTION_LOG] detailing what the agent did, the [ORIGINAL_USER_REQUEST] for context, and a set of [TRIAGE_CATEGORIES] and [PRIORITY_LEVELS] that match your operational taxonomy. The output is a strict JSON package containing a priority classification, category tags, a human-readable summary, and a list of pending decisions. This structure is critical because it allows the downstream application to parse the triage result and automatically route the task to the correct queue, assign it to an on-call specialist, or populate a review dashboard without a human manually reading a wall of text.

Do not use this prompt for tasks where the agent's reasoning is the primary artifact for review, such as debugging an agent's faulty logic. For that, use a dedicated 'Handoff Prompt with Reasoning Trace.' Avoid this template when the task is trivial and the cost of human review outweighs the risk of an error; a simple confidence-based escalation prompt is a better fit. Before implementing, define clear evaluation criteria: the priority and category must match a human expert's judgment in a blind test, and the summary must contain every piece of information a human needs to act without consulting the raw agent log. A common failure mode is the model hallucinating a resolution status or injecting a recommendation that the agent never made, so your harness must validate that the output strictly derives from the provided [AGENT_ACTION_LOG].

The following prompt template is designed to be copied directly into your application or orchestration layer. It instructs a model to act as a triage agent, transforming raw agent logs, task context, and outputs into a structured summary suitable for a human reviewer. The goal is to eliminate the need for a human to read through verbose agent traces by providing a classified, prioritized, and context-rich handoff package. All dynamic inputs are represented as square-bracket placeholders, which your application must populate before sending the request to the model.

textCopy
You are a Triage Agent responsible for preparing a task for human review. Your output will be read by a human operator who has no prior context on this task. You must be concise, accurate, and structured.

Analyze the provided agent task context, actions taken, and final output. Classify the task and produce a structured triage summary.

## INPUTS
- **Task Context:** [TASK_CONTEXT]
- **Agent Actions Log:** [AGENT_ACTIONS_LOG]
- **Agent Final Output:** [AGENT_FINAL_OUTPUT]
- **Escalation Reason:** [ESCALATION_REASON]

## CONSTRAINTS
- Do not invent information not present in the inputs.
- If the agent's output contains a confidence score, include it. Otherwise, do not guess.
- Flag any missing critical information as "Unknown".
- The priority must be derived from the escalation reason and task context, not assumed.

## OUTPUT_SCHEMA
Respond with a single JSON object conforming to this structure:
{
  "triage_summary": {
    "task_id": "string (from context)",
    "priority": "CRITICAL | HIGH | MEDIUM | LOW",
    "category_tags": ["string", "..."],
    "one_line_summary": "string (max 150 chars)",
    "escalation_trigger": "string (the specific event or condition that caused the handoff)",
    "agent_actions_summary": ["string (concise action description)", "..."],
    "pending_decisions": ["string (explicit question or decision needed from human)", "..."],
    "key_evidence": [
      {
        "claim": "string",
        "source": "string (reference to agent log or context)"
      }
    ],
    "risk_assessment": "string (brief assessment of risk if not handled promptly)",
    "recommended_next_step": "string (actionable suggestion for the human reviewer)"
  }
}

## EXAMPLES
- If the agent failed due to a tool timeout, the escalation_trigger is "Tool timeout after 3 retries" and priority might be HIGH.
- If the agent completed but has low confidence, the escalation_trigger is "Low confidence score (0.45) on final output" and pending_decisions should ask the human to verify the output.

To adapt this template for your specific workflow, replace the placeholders in the INPUTS section with data from your agent execution environment. [TASK_CONTEXT] should contain the original user request and any relevant session state. [AGENT_ACTIONS_LOG] should be a structured summary of tool calls, reasoning steps, and errors, not a raw, unformatted dump. The [OUTPUT_SCHEMA] can be extended with domain-specific fields, but avoid removing the core triage fields like priority, pending_decisions, and key_evidence, as these are essential for a human to pick up the task efficiently. Always validate the model's JSON output against this schema before presenting it to a human reviewer; a malformed triage object is as useless as no triage at all.

The Human-in-the-Loop Triage Prompt Template is designed to sit at the boundary between an autonomous agent system and a human review queue. In a typical implementation, an agent completes its work, reaches a decision point or confidence threshold, and calls this prompt to produce a structured triage summary. The application layer then ingests the structured output, enriches it with metadata (agent ID, session ID, timestamp), and pushes it to a review queue—such as a ticketing system, a Slack channel, or a custom review dashboard. The prompt is not the handoff; it is the formatter that makes the handoff legible and actionable for a human who was not watching the agent work.

The harness must validate the output against a strict schema before routing to a human. At minimum, confirm that priority is one of the allowed enum values (e.g., low, medium, high, critical), that category_tags is a non-empty array of strings, and that the summary field contains enough substance to be useful (a length floor of 50 characters is a reasonable starting point). If the output fails validation, retry once with a repair prompt that includes the schema violations. If the retry also fails, escalate the raw agent context to a human with a flag indicating the triage prompt failed—never silently drop a handoff. For high-risk domains, add a mandatory human approval step before the triage summary is considered final; the reviewer should be able to correct the priority or category before the task is assigned.

Model choice matters here. The triage prompt requires reliable structured output and consistent classification, so prefer models with strong instruction-following and JSON mode support (e.g., GPT-4o, Claude 3.5 Sonnet). Avoid models that frequently drop fields or hallucinate enum values. If you are using a model that does not support strict JSON mode, wrap the output in a repair step that extracts and normalizes the JSON from the raw response. Log every triage output—both successful and failed—along with the agent context that produced it. These logs become your evaluation dataset for measuring classification accuracy, priority drift, and whether the triage summaries actually help human reviewers make faster, better decisions. The next step after implementing the harness is to run a calibration exercise: have two human reviewers independently triage the same set of agent outputs, compare their classifications to the prompt's output, and tune the prompt's few-shot examples and priority definitions until agreement is high enough for your operational risk tolerance.