Grill

Use this skill only when the user explicitly asks for grilling plus durable documentation or directly invokes this skill; the agents coordinator handing it an unclear goal, or one decision needing the user's confirmation, before its plan is a direct invocation (its question is the user's judgement — target, measure, scope — that repository facts shape but never answer), and there the durable record is the project's own PROJECT.md or library/ file, never a repository doc. Complexity, ambiguity, or a possible need for docs alone never activates it. This skill carries its own interview and documentation contract. Do not load or invoke domain-modeling at runtime.

Interview Contract

1. Research discoverable facts in the repository, issue, docs, and code before asking the user. Ask only for judgments or facts that cannot be discovered.
2. Ask one decision-forcing question at a time. State the recommended answer and the reason briefly, then wait for the answer.
3. Ask every interview question in plain text as an ordinary assistant response. Do not use a question tool.
   • When a bounded decision benefits from 2-3 short, mutually exclusive choices, list them in plain text with the recommended answer first.
   • Invite the user to choose, modify, or discuss the choices instead of forcing a structured selection.
4. Use comparison tables only when the user explicitly requested one or the question concerns agent-product behavior, such as Claude Code versus Codex.
5. Follow dependent decisions until the skill decides the decision tree is exhausted. Never ask a final "are we done?" meta-question.
6. Summarize the settled contract: goals, non-goals, decisions, constraints, risks, validation, and unresolved items.

Background formal evidence