Grilling (la primitive d'interview)
The primitive behind /grill-me and /grill-with-docs, exposed on its own. It models your decision as a tree: each decision opens the ones hanging off it. Work happens in rounds. The frontier is the decisions whose prerequisites are already settled, so the questions askable now without guessing an answer nobody gave yet. The whole frontier goes out in one round, numbered, each question carrying the agent's recommended answer. Central rule: finding facts is the agent's job, never yours, and a question needing an environment fact goes to a subagent instead of being put to you. The decisions stay yours. The session ends when the frontier is empty.
Strengths
- Facts are the agent's job: a question depending on the environment goes to a subagent, it is not put to you
- Frontier recomputed each round: no question rests on an answer nobody has given yet
- Every question arrives with a recommended answer, so you approve fast and argue only the real trade-offs
- An explicit end criterion (empty frontier) instead of the vague feeling of having discussed enough
Limitations
- Stateless and file-free: the bare primitive keeps nothing, everything settled stays in the context window
- A wide frontier round can fire ten questions at once, which is brutal on a poorly framed subject
- Redundant if you already use /grill-me or /grill-with-docs, which drive it internally
Best for
- Product or technical decisions where you feel something stayed implicit without knowing what
- Skill authors who want to reuse the interview loop inside their own workflow
- Short scoping sessions where you want the raw interview, with no CONTEXT.md and no artifact