grill-me Skill Review: What a Claude Code Planning Interview Actually Costs

Matt Pocock's skills repo crossed 234,000 stars this month, and the skill he calls his most popular is 1,987 bytes of markdown. No scripts, no dependencies, no MCP server. It just makes the agent stop and interrogate you before it touches your code.

The pitch is easy to believe. The part nobody publishes is what it costs. Skills are prompt injection by design: the file goes into context, and a skill that works in rounds re-sends its own instructions on every turn. So we installed it, drove a real four-round session, and logged every token.

What the skill actually says

Worth knowing before you install: grill-me isn't the skill. It's a shim. The whole file is a frontmatter block and one line telling the agent to call grilling instead. Same for grill-with-docs at 247 bytes. The real instruction lives in skills/productivity/grilling/SKILL.md, and it's built around one idea:

Then it defines a frontier: the set of questions whose prerequisites are already settled. Ask the entire frontier in one round, numbered, each with a recommended answer. Wait. The user's answers push the frontier outward and unblock the questions that depended on them. Recompute, ask again. Session ends when the frontier is empty.

Two rules in there do most of the work. First, facts are the agent's job, decisions are yours — if a question needs something from the filesystem, the agent dispatches a sub-agent instead of asking you. Second, a question whose answer depends on another open question belongs to a later round. That's what stops the usual wall of twenty shotgun questions where half depend on the other half.

The measured run

We ran claude-sonnet-5 through an OpenAI-compatible endpoint with usage accounting on, driving a realistic prompt: add a public REST API to an existing Next.js SaaS app so customers can pull usage data. Four rounds, with substantive answers each time, until the agent declared the frontier empty.

One methodology note, because it changes the numbers a lot. Our test client adds a fixed system preamble to every request. We measured it with a two-token control prompt before the run — 6,479 prompt tokens for the word "ok" — and subtracted that baseline from every turn. Every figure below is net of that overhead. If you skip this step you overstate a four-round skill session by about 26,000 tokens, which is where inflated skill-cost claims tend to come from.

Round one is the expensive one and it should be: that's where the skill file loads and the agent produces the full opening frontier. It gave five numbered questions, each with a recommendation — consumer type, auth mechanism, scope of the usage data, versioning, rate limiting. It also announced it was dispatching a sub-agent to check the router version, existing auth, ORM, and tenant model, exactly as the skill instructs, and explicitly said it wasn't blocking the rest of the round on that.

Rounds two through four collapse fast in output while input keeps climbing. That's the shape of any multi-turn skill: history re-sends every turn, so input grows while the agent has less left to say. Round two was the peak input at 3,957 net tokens.

Cost across models

Same 13,519 net tokens, priced against current published rates. Anthropic confirmed Claude Sonnet 5 stays at $2 and $10 per million tokens — the increase to $3 and $15 that was scheduled for September 1, 2026 was cancelled.

Nine cents on Opus 5 for a planning session. Worth saying plainly: the skill is not the expensive part of your day.

Compared to just asking

We ran the identical feature request with no skill loaded — plain "give me an implementation plan." That's the honest baseline, because the alternative to a grilling session isn't silence, it's an unguided prompt.

A 37x cost multiple that is still under four cents. The comparison that matters isn't skill versus no skill, it's four cents versus the cost of the agent confidently building the wrong auth model because nobody asked whether the API was server-to-server or browser-facing.

In our unguided run the agent went straight to reading package.json and never raised the consumer-type question at all. With the skill, that was question one.

How to install it

Two routes, and the repo is clear that you should pick one. The plugin gives you a managed read-only bundle that updates when the author ships. The installer copies editable files into your project so you can hack on them.

# Claude Code, managed bundle, auto-updates
claude plugins install mattpocock-skills

# Codex and other agents, or editable local copies
npx skills@latest add mattpocock/skills

Install both and you get every skill twice. If you use the npx route, include setup-matt-pocock-skills in your selection — several skills read config it writes, and code-review will refuse to run without docs/agents/issue-tracker.md. Then run /setup-matt-pocock-skills once per repo to pick your issue tracker, triage labels, and docs location.

Invoke it with /grill-me for non-code work, or /grill-with-docs when you want the docs-aware variant. Note that grill-me sets disable-model-invocation: true in its frontmatter, so the agent won't fire it on its own. You ask for it.

Where it earns its keep, and where it doesn't

Use it when the cost of a wrong assumption is measured in hours: new public interfaces, auth and permission models, data migrations, anything where "I thought you meant X" means rewriting the afternoon. Those are the cases where five questions up front beat five hundred lines of wrong code.

Skip it for one-line fixes, mechanical refactors, and anything you've already specced in detail. The skill is built to keep going until the frontier is empty, and on a trivial task that's just rounds of confirmation. There's also a real ergonomic cost: it will not act until you confirm shared understanding, which is the whole point and also genuinely annoying when you're in a hurry and already know what you want.

One practical note on model choice. The expensive round is the first one, where the agent has to reason about your whole design tree. Cheap models tend to produce flatter, more generic frontiers. Running round one on a stronger model and later confirmation rounds on a cheaper one is a reasonable split, and at these token counts you can afford to just use the good model throughout.

FAQ

What does the grill-me skill do?

It makes a coding agent interview you in rounds before writing code. Your feature is mapped as a design tree, and each round asks every question whose prerequisites are settled, with the agent's recommended answer attached. grill-me is a 157-byte shim; the real logic is the 1,987-byte grilling skill.

How much does a grill-me session cost in tokens?

Our four-round session used 12,302 net input tokens and 1,217 output tokens, 13,519 total. That's $0.037 on Claude Sonnet 5, $0.092 on Claude Opus 5, $0.018 on Claude Haiku 4.5.

Is the grilling skill safe to install?

The skills in this repo are plain markdown with no executable code, and the repository is MIT licensed. We read the files before running them. That said, treat every third-party skill as untrusted input and read it yourself — a skill is instructions handed to an agent that has your tools.

Does it work outside Claude Code?

Yes. The skill is model-agnostic markdown, and the repo ships an agents/openai.yaml alongside the grilling skill for other harnesses. The npx skills installer targets multiple coding agents.

Skill by Matt Pocock, MIT licensed. Source: github.com/mattpocock/skills. Quoted lines are short excerpts for review purposes. Measurements taken August 24, 2026 against claude-sonnet-5; your token counts will vary with prompt and codebase.