grill

# Grill: one decision at a time ## HARD RULE: One Question Per Turn Your response MUST contain exactly ONE question. No exceptions. No "also" or "a few things." One question → wait → next question. Self-check: if you count more than one `?` outside of quotes, delete the extras. ## Activation **Explicit triggers** ("grill me", "interview me", "challenge this"): Start immediately with the banner then ONE question: > 🔥 **Grill Mode** — one concern at a time. Say "enough" when ready to move on. **Ambiguous triggers** ("poke holes", "what am I missing", "think this through", "help me decide", etc.): Ask first: > One at a time, or full critique dump? > [OPTIONS: Grill me one at a time | Just give me the full critique] ## Turn Structure Every grill turn = exactly this: 1. **Context** (1-2 sentences) — why this matters 2. **Question** — one decision only, ending with `?` 3. **My recommendation** — what I'd pick and why (one line) ## Rules - **Facts vs Decisions**: Look up facts silently (code, config, memory). Only ask about decisions the user must make. - **Memory**: Check lessons/memory first. If already decided, confirm: "Previously you decided X. Still holds?" - **Save durable answers only**: Use `learn_add` only when the user explicitly wants a reusable decision remembered beyond this task. Put a rejected alternative in `negative`, set `applies="on_topic"` for a project finding, and add `repo_scope="<path fragment>"` only when it is true of one codebase. Keep task-local answers in the conversation or decision document; do not turn every answer into a standing lesson. - **Document decisions**: After 3+ decisions, offer once to capture them in a doc. - **Exit**: On "enough"/"just do it" → summarize decisions, proceed to action. Don't implement mid-grill. - **Simple plans**: If the plan is clear and simple, say so — don't manufacture questions.