If .claude/skills/grill-me/SKILL.md.disabled exists, stop before reading a fallback.
Project configuration: If the current project contains .claude/skills/grill-me/SKILL.md, read and apply it instead of this bundled default. The project copy is the capability source of truth.
Repository context: Read ../setup/references/task-context.md when resolving a task, project commands, named agent roles, or legacy .claude/ resource paths. Use repository evidence and applicable project instructions; a missing local workflow copy does not require setup. Resolve bundled resources from the installed skill, never from the target cwd.
Grill Me
Pressure-test a plan or design until the remaining ambiguity is explicit, bounded, and easy to communicate.
For task work, resolve the reviewed artifact with
../setup/references/task-context.md. Read the linked
entrypoint and relevant sections only; return the compact decision summary to the caller, which
owns any durable task update.
Core Behavior
- First, enumerate every candidate risk you can see across all categories
below (scope, states/flows, assumptions, wording, decision branches) — keep
this internal list so nothing drops silently. A later step filters to what
is worth asking.
- Then surface one focused question at a time, ordered by what most changes
the feature shape if answered wrong — the user must be able to answer
without re-reading the whole thread; batched questions get half-answered
or ignored.
- For each question, provide your recommended answer or default position.
- If a question can be answered from the codebase or docs, explore there
instead of asking the user.
- Prefer exposing hidden assumptions over inventing extra scope — the goal
is to sharpen an existing plan, not to bolt on new features under the
guise of stress-testing.
Adversarial Discipline
- When you offer a recommended answer, first spend one beat trying to disprove it. Confirmation is the failure mode — a default you only defended is weaker than one you tried to break.
- Bound the loop. If two consecutive rounds surface only minor or already-considered points, stop and report — continuing is grill theater (motion without new information), not rigor.
- For a high-stakes or irreversible decision, offer the user a fresh-context adversarial check: a subagent prompted to find what's wrong, given the artifact and its constraints — not your preferred answer. Handing over your conclusion biases the reviewer toward agreement. Never invoke an external CLI for this without explicit user authorization.
When Invoked From Discovery
Scan every section of the discovery document (scope, flows, states,
constraints, out-of-scope, open questions, risks) — not only the first
section or headline. Then prioritize:
- Unclear scope boundaries
- Missing states, flows, or edge cases
- Hidden assumptions
- Ambiguous wording a new reader could misinterpret
- Decision branches that materially change the feature shape
Avoid:
- Deep implementation detail unless it changes scope, user experience, risk, or constraints
- Speculative product expansion that should instead be captured as out of scope
Stop Condition
Stop when:
- The main branches of the decision tree are resolved or explicitly cut
- The remaining ambiguity is minor and non-blocking
- The caller can clearly document flow, scope, constraints, and risks
Return To Caller
When the stop condition is met, finish by returning this compact summary
(so the caller — typically /nf or /ct — can paste it into the doc
without editing):
- Clarifications made
- Scope cuts or out-of-scope decisions
- Hidden assumptions uncovered
- Wording fixes or ambiguity reductions
- Remaining risks or blockers
Return this summary after the stop condition is met. Keep the review read-only unless the
caller explicitly authorizes editing the reviewed artifact; a recommendation or a completed
question round is not itself permission to change the plan.
1---2name: grill-me-43description: Use when a plan, design, or discovery document needs stress-testing for hidden assumptions, scope gaps, unresolved4 branches, or ambiguous wording; also when the user explicitly says "grill me".5compatibility: Portable clients may not enforce Claude Code invocation guards or project setup behavior.6metadata:7 source: https://skillmd.com/skills/alexandrbasis/grill-me-48---910<!-- claudops-build: project-config-pointer -->11> If `.claude/skills/grill-me/SKILL.md.disabled` exists, stop before reading a fallback.12> **Project configuration:** If the current project contains `.claude/skills/grill-me/SKILL.md`, read and apply it instead of this bundled default. The project copy is the capability source of truth.13> **Repository context:** Read `../setup/references/task-context.md` when resolving a task, project commands, named agent roles, or legacy `.claude/` resource paths. Use repository evidence and applicable project instructions; a missing local workflow copy does not require setup. Resolve bundled resources from the installed skill, never from the target cwd.1415# Grill Me1617Pressure-test a plan or design until the remaining ambiguity is explicit, bounded, and easy to communicate.1819For task work, resolve the reviewed artifact with20[`../setup/references/task-context.md`](../setup/references/task-context.md). Read the linked21entrypoint and relevant sections only; return the compact decision summary to the caller, which22owns any durable task update.2324## Core Behavior2526- First, enumerate every candidate risk you can see across all categories27 below (scope, states/flows, assumptions, wording, decision branches) — keep28 this internal list so nothing drops silently. A later step filters to what29 is worth asking.30- Then surface one focused question at a time, ordered by what most changes31 the feature shape if answered wrong — the user must be able to answer32 without re-reading the whole thread; batched questions get half-answered33 or ignored.34- For each question, provide your recommended answer or default position.35- If a question can be answered from the codebase or docs, explore there36 instead of asking the user.37- Prefer exposing hidden assumptions over inventing extra scope — the goal38 is to sharpen an existing plan, not to bolt on new features under the39 guise of stress-testing.4041## Adversarial Discipline4243- When you offer a recommended answer, first spend one beat trying to **disprove** it. Confirmation is the failure mode — a default you only defended is weaker than one you tried to break.44- **Bound the loop.** If two consecutive rounds surface only minor or already-considered points, stop and report — continuing is grill theater (motion without new information), not rigor.45- For a high-stakes or irreversible decision, offer the user a **fresh-context adversarial check**: a subagent prompted to *find what's wrong*, given the artifact and its constraints — **not** your preferred answer. Handing over your conclusion biases the reviewer toward agreement. Never invoke an external CLI for this without explicit user authorization.4647## When Invoked From Discovery4849Scan every section of the discovery document (scope, flows, states,50constraints, out-of-scope, open questions, risks) — not only the first51section or headline. Then prioritize:5253- Unclear scope boundaries54- Missing states, flows, or edge cases55- Hidden assumptions56- Ambiguous wording a new reader could misinterpret57- Decision branches that materially change the feature shape5859Avoid:60- Deep implementation detail unless it changes scope, user experience, risk, or constraints61- Speculative product expansion that should instead be captured as out of scope6263## Stop Condition6465Stop when:66- The main branches of the decision tree are resolved or explicitly cut67- The remaining ambiguity is minor and non-blocking68- The caller can clearly document flow, scope, constraints, and risks6970## Return To Caller7172When the stop condition is met, finish by returning this compact summary73(so the caller — typically `/nf` or `/ct` — can paste it into the doc74without editing):7576- Clarifications made77- Scope cuts or out-of-scope decisions78- Hidden assumptions uncovered79- Wording fixes or ambiguity reductions80- Remaining risks or blockers8182Return this summary after the stop condition is met. Keep the review read-only unless the83caller explicitly authorizes editing the reviewed artifact; a recommendation or a completed84question round is not itself permission to change the plan.