Claude Code grill-me: Resolve Open Design Decisions Before You Code

Claude Code grill-me is the entry point that closes unresolved branches of a design tree before you begin implementation or write a plan. The interview loop itself is the model-invoked grilling skill. Community write-ups that still say “three sentences, one question at a time” describe an earlier version. The current default asks frontier questions in rounds.

📑Table of Contents

1. How grill-me, grilling, and grill-with-docs split roles
2. Install with the plugin or npx, not both
3. How round questions and recommended answers work
4. Plan mode: official off versus community pairing
5. What to grill, what to skip, and the next steps
6. FAQ
7. Summary

If you implement from a vague request, undecided branches surface mid-work and force you to rewind. Grilling looks up facts in the repo and leaves decisions to you. It is not a new AI product layer. It is a procedure for naming undecided branches before you write code.

Decide the working order

Decide in this order:

1. The install path
2. The current question format
3. The plan-mode conflict
4. The stop conditions

---

How grill-me, grilling, and grill-with-docs split roles

You invoke a user-facing wrapper. The questions come from grilling. Persist changes into the repo only with grill-with-docs.

The mattpocock/skills README treats misalignment as failure mode #1 and puts a grilling session as the fix. User-invoked skills orchestrate. Model-invoked skills are the reusable discipline. A user-invoked skill must not call another user-invoked skill.

The live grill-me SKILL.md (retrieved 2026-09-01) sets disable-model-invocation: true and is one line: Call the Skill tool with "grilling". grill-me does nothing without grilling installed. grill-with-docs also needs domain-modeling.

How the four skills split

Skill	Invocation	Writes	Fits when
grilling	model-invoked; /grilling also works	none (interview only)	shared loop called by other skills
grill-me	user-invoked; one-line caller of grilling	none; stateless	no repo needed, including non-code decisions
grill-with-docs	user-invoked	CONTEXT.md and ADRs	you have a working directory and want terms plus decisions kept
wayfinder	user-invoked	decision tickets on a map	the work will not fit one session

---

Sources versus early write-ups

Sources (as of September 2026):

• mattpocock/skills README
• grill-me docs
• grill-me SKILL.md

---

What “three sentences” write-ups miss

Write-ups such as Skillselion or Frank The Programmer that treat “three sentences” as the live body do not match the current file. Catalog install counts are a discovery signal, not quality evidence.

Installing “grilling” alone is not enough. Confirm that both the entry skill and the primitive load on the first round. Authoring the SKILL.md itself is covered in ⟦DEVGENT_IMMUTABLE_0⟧.

---

Install with the plugin or npx, not both

Use the official marketplace plugin for Claude Code. Use npx for other agents or an editable copy. Installing both duplicates every skill.

The README path is:

1. Pick plugin or npx. Do not pick both.
2. Plugin: claude plugins install mattpocock-skills, or /plugin install mattpocock-skills in a session.
3. Editable copy: npx skills@latest add mattpocock/skills. Include setup-matt-pocock-skills. Update later with npx skills update.
4. Once per repo, run /setup-matt-pocock-skills (issue tracker, triage labels, docs location).
5. Personal-only skills use ~/.claude/skills/<name>/SKILL.md in Anthropic’s Claude Code Skills docs.

Load rules and install caution

Claude Code loads a SKILL.md only when the skill is used. Invoke with /skill-name, or let a matching description auto-load. Bundled commands such as /doctor are a different set. grill-me is not an Anthropic bundled skill. It rides the same Skills mechanism as any community SKILL.md.

Third-party SKILL.md files are a prompt-injection surface. Do not install a whole pack unread. If you install a subset, keep grilling. If you use grill-with-docs, keep domain-modeling too.

Pick one path, then check on the first round that /grill-me actually loads grilling. The general Skills layout is in the Claude Code Skills complete guide.

---

How round questions and recommended answers work

The default is not one question at a time. Each round asks the whole frontier whose prerequisites are settled, and each item receives a ➡️ recommended answer.

The grilling docs walk it this way:

Run a round

1. Start a fresh conversation with /grill-me (or /grill-with-docs if you will persist into the working directory).
2. Read numbered ❓ questions and ➡️ recommendations as a round, not as a chat drip.
3. Answer by number (for example 1 yes, 2 option B, 3 no plus a reason). If the recommendation contradicts the question wording, say whether you accept the recommendation.
4. Facts are looked up by a sub-agent. Decisions wait for you. An agent that answers its own decisions has broken the skill.
5. An empty frontier is not permission to implement. Stop after the user confirms shared understanding.

---

Official question ranges

Official ranges: grilling docs treat thirteen questions in about three rounds as typical. grill-me docs treat forty-six questions across four rounds as an ordinary session. Two hundred questions usually means the scope was too large or you are in a dumb zone. AI Hero reports sixteen questions on a course-editor feature and 30–50 on complex work. Treat 16 as a short example and 30–50 as the complex-feature side. There is no official question cap.

---

One-at-a-time opt-out

One-at-a-time is a supported opt-out: add When grilling, ask one question at a time. to global CLAUDE.md. Where that line lives is the personal-file side of how to design CLAUDE.md for personal, team, and monorepo setups. /batch-grill-me was folded into grilling; there is no separate sequential skill.

---

When the session is not grilling

The frontier is model judgment, not a computed graph. Weaker or faster models may collapse into a couple of questions and an outline. A session that dumps every question at once with no recommendations is improvisation, not grilling. If the first round is not numbered questions plus recommendations, grilling did not load. You can add a CLAUDE.md rule that forbids implementation without permission.

---

Plan mode: official off versus community pairing

Official grill-me docs turn plan mode off. Plan mode primes the agent to rush a plan, which is the opposite of staying in inquiry. Claude Code Sessions argues the skill pairs well with read-only plan mode. Treat the official docs as primary. Treat pairing as something you must verify.

Official off versus Sessions pairing

Angle	Official grill-me docs	Claude Code Sessions	What to do
Plan mode	Off; it rushes a plan	Pair it; explore, do not implement	Start with official off. If you keep it, cut it the moment inquiry collapses
Start	Fresh conversation, not on an agent-written plan	Rough plan, then grill me	Write your intent in one sentence first; generate a plan file later
Output	Sharp ideas in your head (no files)	A plan with closed branches	Persist with grill-with-docs or a later /to-spec

---

Sources for the plan-mode split

Sources (as of September 2026):

• grill-me docs
• Claude Code Sessions
• Frank The Programmer

---

Where independent sources agree

Independent sources agree on recommended answers, reading the code first, and closing branches before implementation. They diverge when they treat one-at-a-time as the body of the skill. That description does not separate the current round default from the CLAUDE.md opt-out.

Official docs say not to start on top of a plan. Feeding an existing plan often turns interrogation into ratification. If you leave plan mode on, treat a plan file appearing as session failure.

---

What to grill, what to skip, and the next steps

Use it when talking can settle a branch. Cut ungrillable look-and-feel questions to a prototype. Skip one-line fixes; the interview is overhead.

Decision criteria:

• Use it: feature forks, business calls, writing direction, tacit knowledge that should be named before code
• Use grill-with-docs: terms keep drifting and you want CONTEXT.md / ADRs
• Stop: “how should this screen feel?”, 200-question sessions, endless agreement with no decisions of your own
• Skip: the spec is already closed, one-liners, the model deciding for you and starting to implement

After the session

A later /to-spec in the same conversation is optional. grill-me docs say not to start a fresh spec conversation from scratch.

Next actions:

• ✓ Pick plugin or npx (not both)
• ✓ Confirm grilling loaded: first round uses ❓ / ➡️
• ✓ Try one session with plan mode off, as the official docs require
• ✓ If you need one-at-a-time, add the CLAUDE.md opt-out
• ✓ Persist with grill-with-docs; otherwise stay on grill-me
• ✓ Do not implement before shared understanding is confirmed

---

Sources for use and skip

Sources (as of September 2026):

• grill-me docs
• grilling docs

---

FAQ

Are grill-me and grilling the same skill?

Q. Are grill-me and grilling the same skill?

No. grill-me is the user-invoked entry. The current SKILL.md is one line that calls grilling. The interview loop is grilling. Source: grill-me SKILL.md

---

Does it still ask one question at a time?

Q. Does it still ask one question at a time?

The default is rounds. One-at-a-time is restored with When grilling, ask one question at a time. in CLAUDE.md. Source: grilling docs

---

Should I pair it with Claude Code plan mode?

Q. Should I pair it with Claude Code plan mode?

Official docs turn it off. Sessions recommends pairing. Treat official as primary. If you keep plan mode, treat a plan file as inquiry failure. Source: grill-me docs

---

Can I install both the plugin and npx?

Q. Can I install both the plugin and npx?

No. Skills duplicate. Use the plugin for Claude Code and npx for an editable copy. Source: mattpocock/skills

---

Is this an Anthropic bundled skill?

Q. Is this an Anthropic bundled skill?

No. It is a community SKILL.md / plugin. Only placement and invocation use the official Skills mechanism. Source: Claude Code Skills

---

How many questions should a session take?

Q. How many questions should a session take?

There is no official cap. A few rounds is the working range. Split the scope at ~200 questions. The session ends when the frontier is empty and shared understanding is confirmed. Source: grill-me docs

---

Related articles:

• 13624
• 13580
• 13621

Summary

grill-me is the entry that closes design branches before implementation. The body is grilling’s round interview. Official docs turn plan mode off. Community “one question at a time” is the opt-out, not the current default.

Next:

• ✓ Choose one install path
• ✓ Confirm grilling loaded from the first-round format
• ✓ Persist with grill-with-docs, or stay on grill-me
• ✓ Do not implement before confirmation

What remains uncertain: whether plan-mode pairing reproduces, whether weaker models collapse early into an outline, and a known failure where grill-with-docs does not load grilling. Treat the official files as primary and community early shapes as history.