--- name: grill description: "Settle unclear intent before doing any work, by asking the human numbered questions with recommended answers. Use when the goal or done-check is unclear, the request is ambiguous, or the user says \"grill me\", \"stress-test this plan\", or \"aim higher\" (covers the retired ambition skill)." --- # Grill Goal: reach a shared understanding with the human, then record it. Do not start the work. A **decision** is a choice only the human can make (scope, priority, taste). A **fact** is something the workspace or tools can tell you (file contents, versions, what exists). Facts are your job. ## Steps 1. Read the request. Look in the workspace (README, docs, existing files) for every fact you need. Use `ls`, `cat`, `grep`. Never ask the human a question a file can answer; cite the file instead ("README says X"). Before you offer an option that deletes or replaces something, look at that thing with a tool (`ls`, `cat`) and describe what it holds. 2. List the decisions still open. Keep only those whose earlier decisions are already settled. A question that depends on another open question waits for a later round. 3. Ask one **round**: every question from step 2, numbered, in this format: ``` Q1. : Recommended: ``` If a structured question tool exists, use it. Otherwise ask in plain text. 4. Stop. Write nothing else and change no files. Wait for the answers. 5. Take the answers. Recompute: which decisions are now settled, which new questions opened. Ask the next round the same way. Repeat until no open decisions remain. 6. Summarize the settled decisions in a short list and ask: "Is this our shared understanding?" Wait for a yes. 7. After the yes, write the decisions to the file the human named, or `decisions.md` if none was named. One line per decision, plus the goal and done-check in the first two lines. 8. Do the work only if the human then asks for it. Otherwise stop after the file is written. ## Rules - Never act on the plan (no code, no edits other than the decisions file) before the human confirms. - Do not pad. Five sharp questions beat fifteen vague ones. - If an answer is "you decide", take your recommendation and mark it `assumed` in the decisions file. ## Aim higher (only when asked) Use this when the human asks for the ambitious version of a task ("aim higher", "be more ambitious", "how would an expert do this"). 1. If you cannot tell what the task is for, ask that one question and stop. 2. Read `references/ambitious-version.md` in this skill's directory. 3. Reply in exactly this shape, one line each: ``` Outcome: Altitude: "> Depth: "> Range: "> Needs: Check: Which version should I run? ``` 4. Stop and change no files. A bigger goal never widens approval: the new scope, and any access, deploy, message, or spending it needs, waits for the human's yes. ## Done when - The transcript shows at least one round of numbered questions, each with a `Recommended:` line. - No question asks the human for something the workspace already states. - No source or project file was changed before the human confirmed. - After confirmation, the decisions file exists with the goal and done-check on top. - For an aim-higher request: the reply has the `Outcome:` and `Check:` lines, ends with the question, and no file changed.