grill-me

Grill Me

Stress-test the request until its consequential decisions fit together across the affected flow. Research facts, question weak answers, and carry concrete contracts and honest remaining gaps into the execution context.

Read when

• Settling a touched user flow or visible action? Open user-experience.md before proposing choices. Infer routine details from the outcome and existing product; ask only material tradeoffs. Backend-only work skips this.
• About to recommend an answer? Open code-quality.md and code-structure.md. Skip them and you recommend the option the rules forbid.
• Every run: open doctrine.md, the execution context, and the Asking the user and Plain language sections of writing-style.md. Skip them and you mix Questions with Locked in, or ask what the repo already answers.

Process

1. Establish or refresh a compact in-chat execution context from the ask and rediscovered repository, ticket, PR, or diff facts. If a parent already supplied an outcome, current slice, non-goals, area, and ticket/PR, reuse that brief. Re-announce only facts or user decisions that changed or were missing. Before testing solution choices, show the doctrine's intent restatement in your own words, one short plain-English bullet per idea. Reuse a parent's current restatement when it already meets that bar.
2. Separate researched facts, settled decisions, ordinary implementer choices, and unresolved consequential tradeoffs using the doctrine's materiality bar. Research factual gaps and reuse the parent's decisions. Derive ordinary implementation details from the evidence and applicable quality and structure rules; do not turn them into an interview.
3. When the parent is /write-ticket, focus on its ticket-preparation topics. Surface consequential choices affecting behavior, data, contracts, delivery boundaries, dependencies, and authority. The parent derives child count, order, and file lanes from the settled choices.
4. Map choices and their prerequisites using the doctrine's decision rounds. If consequential user-owned choices remain, send a Questions-only batch (no Locked heading). Each question names one concrete choice, the relevant evidence or uncertainty, realistic alternatives, a recommendation with its reason, and practical consequences. Follow decision dependencies and ask the independent choices that can be answered now. Wait for the reply. If everything is settled, omit Questions.
5. Put answers, their reasons, rules that must stay true, corrections, and researched ownership paths directly in the execution context. Ask another batch for an incomplete answer, contradiction, or newly exposed consequential tradeoff. Apply answer pressure; a topic mentioned or recommendation offered is not a settled decision. Missing template categories do not justify another interview.
6. Apply the readiness check to the affected flow. Honor pause, defer, and stop without claiming unresolved work is ready. When material Questions are settled, announce Locked in (tell me if this is wrong) in a separate announce-only message. Include meaningful non-goals, derived split, and shared understanding. Put the corrected line-by-line intent restatement there, keeping current decisions and their reasons and dropping superseded ones. Reuse a current parent lock that already covers these decisions.
7. Issue plans only after the Locked in message stands and every relevant rule has an enforcement and verification owner.

Keep language, choices, rules, and progress in the chat. If the user wants a durable artifact, ask for or honor an approved destination under the shared optional-persistence rule.

If a parent already owns the next step

Follow nested capabilities. Return only changed decisions, their reasons, and remaining material gaps to the parent. It owns the next step; /write-ticket finishes the ticket, and the selected execution skill continues its current lifecycle.

If this is a user one-off

Stop after shared understanding unless the user explicitly asks for the next step. For an authorized build use the Skills table in AGENTS.md:

• Structure needed: /analyze, then the selected execution skill.
• Ready to build: /task-with-tests by default; /task for an explicit command or no-tests request. Carry settled decisions and test refusals.