grill-me

grill-me

Use this as the optional discovery / pre-planning phase inside the atelier workflow.

Interview the user relentlessly about every material aspect of the plan, design, or requested change until the decision tree is resolved enough for plan-task to write a strong execution contract.

Read first

• ../references/task-packet-contract.md
• ../references/persistent-artifacts-contract.md
• ../references/compatibility-policy.md
• ../references/ownership-and-reuse-policy.md
• ../references/final-state-authoring-policy.md
• relevant tracked AGENTS.md chain
• applicable AGENTS.override.md only if local execution constraints matter

Role

This skill owns clarification, not execution.

Keep the main interview context lean and decision-oriented. If clarification requires repo/codebase exploration, spawn or reuse a sidecar teammate named researcher instead of doing the digging yourself. Use model openai-codex/gpt-5.4 by default and choose thinking to match task complexity.

It is for:

• resolving ambiguity before planning
• surfacing hidden constraints
• narrowing decision branches
• testing whether the proposed direction actually makes sense against the codebase
• producing a crisp enough brief that plan-task can write a self-sufficient PLAN.md

It is not for:

• writing implementation code
• doing final plan review
• inventing speculative scope the user did not ask for
• preserving compatibility by default

Interview workflow

1. Understand the current ask.
   • Restate the current design/problem in your own head.
   • Identify what is still ambiguous, missing, or risky.
2. Explore before asking when possible.
   • If the answer is in code, architecture, existing artifacts, tracked AGENTS.md, or nearby boundaries, investigate first.
   • Route non-trivial exploration through a sidecar teammate named researcher; do not load the main interview context with raw repo details.
   • researcher defaults to model openai-codex/gpt-5.4; choose thinking by complexity: minimal/low for quick lookups, medium for bounded multi-file tracing, high/xhigh for ambiguous or cross-cutting investigation.
   • Pull back only the distilled findings needed for the next interview question or brief update.
   • Use questioning only for information that is genuinely missing, preference-driven, or decision-driven.
3. Ask exactly one question at a time.
   • Never batch multiple unrelated questions into one message.
   • Resolve the current branch before moving to the next one.
4. Use multiple-choice format by default.
   • Offer 2–5 concrete answer options.
   • Include Other when the space is open-ended.
   • Make the options mutually exclusive when possible.
5. Mark a recommended option only when confidence is high.
   • Format clearly, e.g. Recommended: B.
   • Include one short explanation of why it is recommended.
   • If confidence is not high, do not force a recommendation.
6. Walk the decision tree top-down.
   • Resolve goals before mechanics.
   • Resolve boundaries before implementation details.
   • Resolve ownership/reuse before new abstractions.
   • Resolve compatibility only if a real external boundary is involved.
7. Stop once the plan can be written cleanly.
   • When the major branches are resolved, summarize the clarified brief or update BRIEF.md if a task packet is already in play.
   • Write the brief as the current clarified model, not as a chronology of how the conversation wandered there.