Skill · Research
grill-the-plan
Interviews you about a plan one question at a time, with a recommended answer for each, until every design decision is resolved, then writes a decision log and a one-page plan. Use when someone says 'grill me', 'poke holes in this plan', 'stress-test my idea', 'what am I missing', 'ask me questions first', or before starting a large build or feature.
The skill itself. Agents read the name and description first and load the rest only when a task matches.
grill-the-plan/SKILL.md
---
name: grill-the-plan
description: Interviews you about a plan one question at a time, with a recommended answer for each, until every design decision is resolved, then writes a decision log and a one-page plan. Use when someone says 'grill me', 'poke holes in this plan', 'stress-test my idea', 'what am I missing', 'ask me questions first', or before starting a large build or feature.
---
# Grill the plan
Interview the user about a plan until every design decision has an answer. No code or build work starts during the interview.
## Steps
1. If the user hasn't described the plan, ask for it in one message: what they want to build or change, and for whom. Then restate it in two sentences and ask them to confirm or correct it.
2. If you're in a repo or have docs, read the relevant code and docs first. Don't ask anything the code or docs already answer. State what you found ("Auth already uses sessions in src/auth/") and move on.
3. Build a decision tree with these branches: goal, users, scope, non-goals, data, edge cases, failure modes, rollout. Add branches this plan needs (pricing, security, migration, cost).
4. Ask one question at a time, starting with the branch that blocks the most other decisions. With each question, give your recommended answer and one line on why, so the user can reply "yes" or correct it.
5. After each answer, show the running list:
- Resolved: branch, decision, one-line reason.
- Open: branches still waiting, in the order you'll ask them.
6. If an answer is vague ("it should be fast", "the usual users"), ask a follow-up that forces a number, a name or an example ("p95 under 300 ms?", "which three customers?"). Don't mark the branch resolved until it has one.
7. When an answer raises new questions, add them as child branches under the branch they came from.
8. Every 10 questions, say how many are still open and ask whether to continue, or to switch to your recommended answers for the low-risk ones.
9. Stop when the open list is empty or the user says stop. If they stop early, list what's still open as risks.
10. Finish with a decision log (every resolved branch, its decision and reason) and a one-page plan: goal, users, scope, non-goals, approach, edge cases handled, failure modes and responses, rollout steps, open risks.
## Rules
- One question per message. Never send a batch.
- Every question comes with a recommended answer.
- Don't write code, files or tickets until the plan is finished and the user says go.
- Don't accept "it depends" or "TBD" as resolved. Keep it open and note what it depends on.
- Point out contradictions right away: quote both answers and ask which one wins.
- Keep questions short. No preamble, no praise.
## Example output
> **Q4 (failure modes):** If the payment webhook arrives twice, what happens?
> Recommended: dedupe on the event ID and ignore repeats, because the provider retries on timeouts.
>
> Resolved: goal, users, scope, non-goals. Open: failure modes, data retention, rollout.