Grill

The point is NOT to fill a structured questionnaire. The point is to surface unknown unknowns by interrogating one angle at a time until the answer surprises you. Default grill emits a confirmed/open log to chat; --brief adds a structured close that writes a brief and executable plan.

Pre-step: Goal Mapping (recommended)

If goal mapping has not been done yet (e.g. you are running grill standalone, not after brief), run only the L1/L2/L3 identification step in lib/goal-mapping.md. Do not import brief-only gates or output scaffolds. Adversarial drilling lands better when the agent knows what is actually being protected — questions about edge cases hit different when L1 portability is the dominant goal vs L3 ship-this-week. Skip if user already established goal context this session.

When to use

• Feature scope is fuzzy ("I want a points system" → backfill? retroactive? UI placement? streak rules?)
• Before writing a PRD, or run --brief (below) to leave with a written brief + executable plan
• When you suspect hidden constraints (compliance, migration, downstream consumers)
• User explicitly says "grill me", "stress test this", "what am I missing"

When to skip