| name | grill-me |
| description | Pressure-test an existing plan, design, or decision, interrogating every decision point one question at a time until hidden assumptions and weak points surface. |
| model | opus |
| effort | high |
Grill Me
Challenge a plan, design, or decision before it is committed. Surface hidden assumptions, weak points, and unresolved dependencies. For generating options instead, use brainstorming.
Reserve it for decisions expensive to reverse: data model changes, public API contracts, security posture, infrastructure choices. Small, low-stakes features skip it.
How to interrogate
- One question at a time, with a recommended answer when one is clear.
- Read the codebase first; ask only what the code cannot answer.
- Navigate the full decision tree, following each answer's branch to its end before the next.
- When an answer reveals a gap or contradiction, name it and ask the follow-up.
Process
- Read
docs/product/vision.md if present. Its Principles and Anti-goals are absolute, so a decision that violates one is a blocking issue, not a trade-off. Start there if any violation exists.
- Derive the decision areas that matter from the artifact itself and the Principles, not a fixed checklist.
- Start with the most load-bearing decision, the one that invalidates the most downstream choices if wrong, and follow its branch. Format each question:
[Topic]: [Question]
Recommended: [recommendation, or "no strong preference"]
- When an answer surfaces a new contradiction, dependency, or principle violation, re-enter step 3 there automatically until nothing new surfaces.
- Summarize: decisions solid, assumptions validated, risks or gaps still open, and what needs deciding before proceeding.
- Ask whether to proceed, revise, or dig deeper. On revision, re-enter at the affected point rather than restarting.
Done When
Every material decision point is covered and the user approves. Suggest implementation-plan to sequence the work.
