← ClaudeAtlas

grilllisted

Relentless one-question-at-a-time interview that stress-tests a plan or design until every branch of the decision tree is resolved and you and the user share one understanding. The reusable alignment primitive the pipeline leans on (analyst wraps it, arch-health runs it, prepare/spike apply it) and a standalone way to interrogate any plan before building. TRIGGER when: the user wants to pressure-test a plan/design, says "grill me" / "interview me" / "poke holes in this" / "what am I missing", OR another skill needs the interview loop. DO NOT TRIGGER when: the user wants a written spec (use /analyst), a full impact analysis (use /prepare), to ALSO capture the terms/decisions as you align (use /grill-with-docs), or simply asked a question to be answered (just answer it).
mik2win/foureyes · ★ 2 · AI & Automation · score 79
Install: claude install-skill mik2win/foureyes
# Grill Interview the user **relentlessly** about every aspect of this plan or design until you reach a shared understanding. Walk down each branch of the decision tree, resolving dependencies between decisions one at a time. The whole point is to close the **alignment gap** — "no-one knows exactly what they want" — *before* code gets written, when changing course is still cheap. This is the invariant interview discipline used across the kit: `/analyst` wraps it to write a spec, `/arch-health` runs it to walk a chosen opportunity, and `/prepare` / `/spike` apply the same one-question discipline when confirming assumptions or framing a hypothesis. You can also run it standalone (`/grill <thing to grill>`) to stress-test any plan — including a non-code one. The subject to grill is `$ARGUMENTS`, the current conversation, or whatever the calling skill hands you. --- ## Phase 0 — Ground (light, optional) Grill runs with or without a project profile — don't block on it (lean on the root `CLAUDE.md` if it carries the facts). - If `.claude/PROJECT.md` exists, read its **Domain** (vocabulary, roles), **Architecture**, and **Plans location** (where a standalone run saves its alignment summary) so your questions use the project's real terms and reference real modules — never invent jargon. - If `CONTEXT.md` (the project glossary, maintained by `/domain-model`) exists, read it too and phrase questions in its **ubiquitous language**. When the user reaches for a term that confl