grill-me

grill-me

Relentlessly interview the user about a plan or design until you both reach shared understanding and every open decision is resolved.

How to run it

• Read the plan, linked docs, code, or local files needed to understand the topic first.
• Ask one question at a time. Wait for the answer before the next question.
• For each question, state your recommended answer and a one-line why.
• Walk each branch of the decision tree. When an answer opens new questions, follow them; resolve dependencies between decisions in order rather than jumping around.
• If a question can be answered by exploring the codebase, explore instead of asking.
• Push on fuzzy words, hidden assumptions, edge cases, dependencies between decisions, and reversibility.
• Keep going until no unresolved branch remains. Then summarize the agreed plan.

Question patterns

• "What outcome would make this change a success?"
• "Which existing concept owns this behavior?"
• "What is the smallest scenario that proves the design works?"
• "What breaks if this assumption is false?"
• "What would a future maintainer be tempted to change back?"
• "Which term should be canonical, and which aliases should be avoided?"

Capturing decisions

Only when the repo already keeps CONTEXT.md, CONTEXT-MAP.md, or docs/adr/, or the interview resolves terminology or a decision worth outliving the conversation. Check for those files before proposing new ones.

• Update the appropriate CONTEXT.md as soon as terminology is resolved — it is a glossary and relationship map for project-specific language, not a spec or scratchpad.
• Use CONTEXT-MAP.md only when the repository has multiple bounded contexts.
• Offer an ADR only for decisions that are hard to reverse, surprising without context, and based on a real tradeoff.
• Create docs lazily: no CONTEXT.md or docs/adr/ until there is real resolved content, and keep them short enough that future agents actually read them.

For glossary format, read CONTEXT-FORMAT.md. For ADR format and thresholds, read ADR-FORMAT.md.

Notes

• Do not soften or batch questions to "be efficient" — the value is the relentlessness. One question, one recommendation, repeat.
• Stop early only if the user explicitly says to stop or accepts all remaining recommendations at once.
• Ported from mattpocock/skills (MIT), adapted to this repo's conventions and voice.