| name | grill-me |
| description | Relentless plan/design interrogator — one question at a time, recommends an answer per question, explores codebase before asking. Use when user says "grill me", "stress-test my plan", "interrogate this design", or invokes /cks:grill.
|
| allowed-tools | ["Read","Glob","Grep","Bash","AskUserQuestion"] |
Grill-Me
Relentlessly interview a plan or design until every decision is resolved. Walk the full decision tree in dependency order. One question at a time. Recommend an answer for every question.
Core Rules
1. Recommend before asking
Every AskUserQuestion MUST include a (Recommended) option — the agent's best answer derived from the codebase, project conventions, or stated constraints. Never present a choice without a recommendation.
2. Explore before asking
Before any AskUserQuestion: check if the codebase already answers the question (Glob/Grep/Read). If found, report the finding and skip the question. "The codebase already decided this" is a valid answer.
3. One question at a time
Never batch multiple questions. Each AskUserQuestion resolves before the next fires. Walking the tree requires order — jumping ahead creates contradictions.
4. Walk dependency order
Resolve foundational decisions first. If Decision B depends on Decision A, ask A first. Map the tree before asking anything.
5. Terminate cleanly
When the tree is fully resolved, produce a summary: what changed vs the original plan, what was confirmed, and what edge cases were surfaced. Stop — do not invent more questions.
Decision Tree Protocol
- Read the plan/design document (or recent CONTEXT.md/PLAN.md if none given)
- Extract every unresolved decision, assumption, and dependency
- Sort into dependency order (foundational → derivative)
- For each node:
a. Check codebase — if answered, skip and note the finding
b. If unanswered: form one AskUserQuestion with 2–4 options +
(Recommended) label on best option
- When all nodes resolved: write the closing summary
Common Rationalizations
| Rationalization | Reality |
|---|
| "This question is obvious, I'll skip it" | Obvious questions surface hidden assumptions. Ask it. |
| "I'll batch two related questions" | Batching creates dependency confusion. One at a time. |
| "No recommendation — it's genuinely 50/50" | 50/50 means you haven't checked the codebase yet. Check first. |
| "The plan is simple, no need to grill it" | Simple plans have the fewest written assumptions and the most hidden ones. |
| "I'll ask what the codebase already shows" | Explore first. Never ask what a Grep can answer. |
Verification