grilling

A model-invoked productivity skill: “Grill the user relentlessly about a plan, decision, or idea.” It is the reusable interview loop that underlies both mp-grill-me and mp-grill-with-docs, and the canonical implementation of pattern-grilling — keep asking until decision branches resolve — factored out so other skills compose it rather than re-implement it.

How it works

• Round by round over the frontier. It maps the decision tree and asks the whole frontier (every question whose prerequisites are already settled) in one numbered round, then recomputes the frontier from the answers and asks the next round. “Same 13 questions land in ~3 rounds instead of 13.” The session ends when the frontier is empty. A global CLAUDE.md line opts back into one-question-at-a-time.
• Facts vs. decisions. Facts the environment can answer are looked up, dispatched to background sub-agents so research never blocks a round; only questions downstream of a running exploration wait for it. Decisions are put to the human and wait for an answer, so when another skill (e.g. mp-wayfinder or mp-triage) runs grilling inside a resolve-the-ticket frame, the agent can’t race ahead and answer its own decisions (the HITL guarantee).
• Confirmation gate. The agent won’t act on it until the user confirms shared understanding has been reached.
• Fixed question shape. Each question is ❓ **Q1** - **<title>**, then the body (prose or multiple choices), then the recommendation on its own ➡️ line; questions in a round are separated by ---. A round reads as a scannable numbered list you can answer by number.

mp-grill-me, mp-grill-with-docs and mp-triage all run the frontier a round at a time. The frontier leading word recurs in mp-implement-spec (the ready tickets in a task graph).

See Also

• mp-grill-me, mp-grill-with-docs — the user-facing skills built on this loop.
• mp-to-questionnaire — the inverse: grills the send when the answers live in someone else’s head.
• stage-align — the canonical stage this implements.

Version history

v1.3.0 (2026-10-04)

• Questions in a round are separated by a horizontal rule (---); em-dashes removed from the skill text.

v1.2.0 (2026-08-05)

• Reworked from one-question-at-a-time to round-by-round over the frontier, with facts dispatched to background sub-agents and a fixed ❓ / ➡️ question shape.
• Reworded for general use: “this plan” → “this”, “exploring the codebase” → “exploring the environment”; the description now covers “a plan, decision, or idea” rather than only a plan.

v1.1.0 (2026-07-08; ingested 2026-07-09)

• Added the confirmation gate and the facts vs. decisions split (the HITL guarantee).

Initial ingest (2026-06-28)

• Paged as the reusable interview loop behind the grill skills, asking one question at a time.