grill-me skill for Claude

A coding agent that builds the wrong thing quickly is still building the wrong thing. Matt Pocock's README names misalignment as the most common failure in software work, and his fix is a grilling session: the agent asks you detailed questions before anyone writes code. grill-me packages that session as a skill. It works for ideas outside software too. This page covers how it runs, how to install it and how to get a useful session out of it.

What the skill does

According to its documentation, grill-me takes a loose idea and interviews you until you can commit to it. You do not need a plan to start, because producing one is the point of the session.

The questioning logic lives in a companion skill called grilling, which grill-me calls. It works like this:

• A design tree. Every decision branches into the decisions that depend on it.
• Questions in rounds. Each round asks every question whose prerequisites you have already answered, numbered, with a recommendation for each. You are never asked something that depends on an answer you have not given.
• Facts vs decisions. The agent looks up facts itself and leaves decisions to you.
• A clear end. The session ends when every branch is visited and you confirm you share an understanding, before any work starts.

grill-me is stateless. It writes no files and leaves nothing behind except a sharper version of the idea. The docs call a session of forty-six questions across four rounds ordinary.

Who publishes it

Matt Pocock, a community author, in his mattpocock/skills repository under the MIT license. The README calls grill-me and its sibling grill-with-docs his most popular skills. Anthropic does not publish it.

Where it works

The repository documents Claude Code, through a plugin, and Codex and other agents, through the Skills CLI. It does not give steps for Claude.ai. Because grill-me only hands off to grilling, a manual upload would need both skills.

How to install it

Claude Code. The README says the plugin is in Claude Code's official marketplace, so there is nothing to add first:

claude plugins install mattpocock-skills

Or, inside a session: /plugin install mattpocock-skills. This installs the whole set as a read-only bundle that updates when the author ships changes.

Codex and other agents, or if you want editable copies in your repository:

npx skills@latest add mattpocock/skills

The README warns against using both methods, which leaves every skill installed twice. It also asks you to run /setup-matt-pocock-skills once per repository; that setup asks which issue tracker you use, which labels you apply when triaging, and where to save docs.

How to run a good session

You start it yourself by typing /grill-me; the skill is set so the agent never starts it on its own. Claude Code lists plugin skills in the / menu under the plugin name, so look for /mattpocock-skills:grill-me if the short form does not appear. The author's docs add a few rules:

• Start in a fresh conversation, not on top of a plan an agent already wrote.
• Leave plan mode off. It pushes the agent to rush toward a plan.
• Steer. Push back, say when the scope drifts, and answer "I don't know" when that is true. A session where you agree with everything produces a plan you nodded at.
• Stop grilling on questions only a prototype can answer, such as how an interaction should feel. Build a rough version, look at it, then answer.
• If you prefer one question at a time, add "When grilling, ask one question at a time." to your global CLAUDE.md.

/grill-me

I want to [the idea, in one or two sentences].
It is for [who will use it or be affected].
What I already know: [constraints, budget, deadline].
What worries me: [the part I am least sure about].

When it helps

• Before a feature, so the agent and you agree on scope.
• Product and business calls: pricing, a launch plan, what to build next. Our business prompts pair well with it.
• Writing projects and personal decisions. The docs say people point it at anything that will not sit still in their head; our self-improvement prompts cover similar ground without a skill.

When you have a codebase to align against, the author recommends grill-with-docs, which reads your code and records terms and decisions in GLOSSARY.md and decision records.

Limits

• Your answers set the quality. The docs warn that passive answers produce confident nonsense.
• Big scope balloons. A session that runs to two hundred questions usually means the work should be split first. Very long sessions also fill the context window, and question quality drops.
• Model choice matters. The author says grilling leans on the model's sense of how systems break, so use your strongest model.

For other skills, see the Claude Skills catalog; for how skills are built, read our Claude Skills manual. Prompts for the build phase are in Claude Code prompts.

FAQ

Is grill-me an official Anthropic skill?

No. It is part of Matt Pocock's open-source skills collection on GitHub, under the MIT license.

Does grill-me work only for code?

No. Its docs say it needs no repository and the subject does not have to be code.

What happens after the session?

Nothing is saved to disk. If the idea is software, the docs suggest continuing in the same conversation with the to-spec skill to write a spec.