mk:grill

What This Skill Does

Turns the usual direction around: instead of Claude answering or proposing, Claude interviews you about a plan or design you already own — one question at a time — and drives every unresolved branch of the decision tree to a resolution. The output is a knowledge document that front-loads the tacit context a downstream build skill would otherwise have to guess at.

It accepts either entry mode:

• A doc path — a plan / design / spec file. grill reads it and seeds the decision tree from its open branches (unresolved choices, unjustified decisions, TODOs, vague nouns, assumptions stated as fact). It never edits the source.
• A verbal idea / topic — grill seeds an empty tree and builds it through questioning.

The signature mechanic is exactly one question per turn — deliberately not a batched multi-question prompt. One sharp question, your answer, an immediate checkpoint, then the tree is re-evaluated and the next highest-leverage branch is asked. Five questions or an hour are both valid lengths; depth, not count, is the target.

When to Use

• "Grill me on this plan"
• "Stress-test my design / interrogate my plan"
• "Get grilled on my design before I build it"
• You have a half-formed plan or design and want every assumption, ambiguity, and unresolved choice surfaced and answered before handing it to a build skill.

When NOT to Use

Need	Use this instead
Propose / compare technical approaches	mk:brainstorming
"Should we build this" — product validation	mk:office-hours
Claude should ANSWER questions about the repo	mk:ask-me
One-shot review of an existing plan	mk:plan-ceo-review
Apply named reasoning lenses to existing findings	mk:elicit
Multi-perspective architecture debate	mk:party
Constitution derived from the CODEBASE (not the human)	mk:project-context
Bug / error / root cause	mk:investigate

Usage

# Grill an existing plan/design doc (read-only on the source)
/mk:grill tasks/plans/260613-1254-payment-retry/plan.md

# Grill a verbal idea — no doc required
/mk:grill I want to add a retry queue for failed payment webhooks

Argument hint: [doc path | verbal idea or topic]

The only output is docs/knowledge/<slug>.md, written and updated after every answer. The source doc, if any, is never modified.

Core Loop

1. Pick the single highest-leverage unresolved branch.
2. Ask exactly one specific, concrete question about it — no batching, no menus.
3. Wait for your answer.
4. Restate the answer in one line to confirm understanding.
5. Checkpoint immediately to docs/knowledge/<slug>.md before asking the next question.
6. Re-evaluate the tree — the answer may close branches, open new ones, or expose a contradiction with an earlier answer (surfaced on the spot).
7. Repeat until termination conditions are met.

Gap Flagging

When you cannot answer, answer thinly, or say "I don't own that", grill does not paper over it. It records an Open Gap with the unresolved question, why it is unresolved, and who to ask — then tells you to go get the detail and drop it back into the knowledge doc. A flagged gap is a valid outcome, not a failure.

Hybrid Termination

At each checkpoint grill surfaces a readiness signal — the remaining open branches plus its judgment of whether high-leverage questions are exhausted. When branches drain, it proposes wrap-up and asks you to confirm "done" or "keep going". It never auto-terminates silently and never loops forever; you hold the stop button.

On confirmed wrap-up it offers three handoffs:

• mk:skill-creator — if the resolved design is a skill to build.
• mk:plan-creator (or mk:brainstorming if approaches are still open) — if it is a feature/system to plan.
• mk:project-context — to merge the extracted human knowledge into the project constitution.

grill offers all three and lets you pick; it does not auto-invoke the downstream skill.

Output Format

docs/knowledge/<slug>.md:

# Knowledge: <Plan/Design Title>

**Source:** <doc path | "verbal idea">
**Started:** <date>  ·  **Status:** in-progress | wrapped

## Decisions
- **<branch>:** <answer> — <why>

## Q&A Log
1. **Q:** <question>
   **A:** <answer>

## Highlights
- <insight>

## Open Branches
- <branch>

## Open Gaps
- **<question>** — <why unresolved> → ask: <who/source>

Safety Boundaries

• Read-only on the source. The plan/design doc is never edited; corrections flow through the chosen handoff. The only file grill writes is docs/knowledge/<slug>.md.
• No AskUserQuestion. The one-question-at-a-time loop is plain prose — batched question prompts are deliberately excluded from the skill's tools.
• Never proposes the answer. The moment a suggestion would replace a user decision, that is mk:brainstorming territory; grill asks, it does not decide.