Grill Me

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.