goal-loop-auditor

name	goal-loop-auditor
description	Audit goal, autonomous loop, research loop, execution loop, review loop, monitoring loop, or long-running agent prompts. Use when the user asks to review, harden, score, debug, or improve a goal/loop prompt for missing facts, ambiguous scope, weak validation, source-quality gaps, output-contract gaps, conflict-handling problems, stop-condition problems, permissions, or hidden assumptions.

Goal Loop Auditor

Evaluate whether a goal/loop prompt is safe to launch. Prioritize defects that would cause wrong work, unverifiable results, scope drift, unsafe action, or guessing.

Workflow

1. Identify loop type: research, execution, review, monitoring, or mixed.
2. Check launch fields: PROJECT, GOAL, MODE, CONTEXT, SCOPE, RULES, SOURCES, METHOD, OUTPUT, ON CONFLICT, STOP CONDITION, DONE CRITERIA.
3. For execution loops, also check PERMISSIONS for writes, dependencies, migrations, commits, pushes, network, secrets, production, and destructive actions.
4. Report only material findings. Group related missing fields instead of listing every absent label separately.
5. Produce a corrected prompt only when assumptions are explicit and non-blocking.

Severity

• Blocking: should not launch until fixed.
• High: likely wrong, unsafe, or unverifiable.
• Medium: likely inefficient or inconsistent.
• Low: polish only.

Missing GOAL, SCOPE, RULES, OUTPUT, or STOP CONDITION is Blocking.

Audit Standard

Check that the prompt has a concrete deliverable, explicit exclusions, observable validation, source rules, output schema, conflict handling, halt cases, no hidden conversation dependency, and checkpoints or budget for long-running work. Fold conflict, checkpoint, and budget gaps into grouped findings unless they are the primary launch risk.

If blocking facts prevent a safe corrected prompt, ask exactly one next blocking question with Recommended answer: instead of listing every missing fact.

Output

Return this shape:

Audit result: [Ready | Needs revision | Do not launch]

Findings:
- [Severity] [Field/group]: [problem]
  Impact: [why it matters]
  Fix: [specific change]

Next blocking question:
[Ask exactly one question when a safe correction needs user input.]
Recommended answer: [default that preserves momentum without inventing concrete facts.]

Assumptions:
- [non-blocking assumptions only]

Corrected prompt:
```text
[include only when safe]
```

Keep findings to the smallest set that changes launch safety. If there are no material issues, return Audit result: Ready and residual risks only.