The grill-me Skill for Claude Code

You have probably hit this one. You type a paragraph describing some feature, Claude reads it, and half a minute later it is off building. The thing it builds answers the prompt you wrote, not the thing you actually had in mind. The requirements were fuzzy in your head, so they came out fuzzy on the keyboard, and the model just filled the gaps with reasonable-looking guesses.

grill-me turns that around. Rather than you briefing Claude so it can start coding, Claude interviews you first. It pulls your plan apart one question at a time until there is nothing left that is vague enough to build wrong. I went in expecting a gimmick and came out using it most weeks.

What grill-me actually is

It is a Claude Code skill, which is just a small SKILL.md file that Claude finds and loads by itself when the moment fits. Matt Pocock wrote it (you may know him from Total TypeScript), and it lives in his public mattpocock/skills repo, which is basically a dump of the skills from his own .claude directory.

The skill is community-made, not something Anthropic ships, and it took off. The repo collected tens of thousands of stars after he posted about it, and grill-me was the one everyone fixated on. Partly that is because it names a problem people run into constantly, and partly because there is almost nothing to it.

The whole skill is three sentences

That last point is worth dwelling on. No scripts, no scaffolding, no tooling. The entire skill is a description and three sentences of instruction:

---
name: grill-me
description: Interview the user relentlessly about a plan or design until reaching shared understanding, resolving each branch of the decision tree. Use when user wants to stress-test a plan, get grilled on their design, or mentions "grill me".
---

Interview me relentlessly about every aspect of this plan until we reach a shared understanding. Walk down each branch of the design tree, resolving dependencies between decisions one-by-one. For each question, provide your recommended answer.

Ask the questions one at a time.

If a question can be answered by exploring the codebase, explore the codebase instead.

That is the whole file, sitting at skills/productivity/grill-me/SKILL.md. All the work is being done by the prompt. The line that carries the weight is "walk down each branch of the design tree." It tells Claude to treat your feature as a tree of decisions, where every choice opens up a few more, instead of one block of text to skim and act on. The other two sentences just keep things sane: ask one question at a time, and go read the code rather than asking when the repo already knows the answer.

What it feels like to use

You kick it off by saying "grill me" (or running /grill-me) next to whatever you want torn apart. Then the roles swap. Claude starts asking, and it does not run out of questions at the convenient point.

A few details make it work:

• One question at a time. You are not handed twenty bullet points to clear in one go. Each question gets its own turn, so you have to actually think about it rather than skim past.
• It brings its own answer. Every question comes with the option Claude would pick and why. Reacting to a concrete proposal is a lot easier than producing one from nothing, which is what keeps the session moving.
• It reads the code first. When the repo can settle a question, Claude goes and looks instead of making you recite something that is already written down.
• It keeps going. It works down the branches until you genuinely agree on the plan, not until the conversation feels long enough.

The annoying questions are the useful ones. "What happens when this fails halfway through?" is exactly the kind of thing you were quietly planning to deal with later. Answering it before any code exists is far cheaper than running into it three files deep.

Installing it

The fastest way is Pocock's installer, which grabs the skills from the repo and lets you choose what to add:

npx skills@latest add mattpocock/skills

If you would rather skip the installer, remember a skill is just a file in a folder. Save the SKILL.md above to ~/.claude/skills/grill-me/SKILL.md and Claude Code will find it on its own, then call it with /grill-me. Every skill works the same way, a SKILL.md sitting in its own folder under ~/.claude/skills/. If you want to see how skills and plugins hang together, I wrote up how to build a Claude Code plugin separately.

Beyond coding, and grill-with-docs

None of this is really about code. Any time your thinking is messier than you want to admit, whether it is a product call, a piece of writing, or an architecture you cannot quite commit to, being walked down the decision tree drags the fuzzy bits into the light. Pocock has said he uses it well outside of programming, which makes sense once you notice it is a thinking tool that happens to be wearing a coding hat.

For codebases there is a follow-up called grill-with-docs. It grills you the same way, but as decisions firm up it writes them down: a CONTEXT.md glossary of your project's terms, plus Architecture Decision Records dropped in as you go. So the grilling stops being a single conversation and starts to pile up into shared context you keep between sessions. There are community forks too (Jekudy/grillme-skill is one), which tends to happen when an idea lands.

Worth it?

For anything bigger than a one-liner, yes. The cheapest tokens you spend are the ones on planning, and the most expensive bug is the one you confidently build from a vague brief. grill-me is a structured way to pay the small cost first. If that is your kind of thing, getting more out of Claude Code goes deeper on prompting and token economy for the same basic reason: sort the plan out before you spend the budget building it.