/grill-me from Matt Pocock's skills repo makes your coding agent question the plan first

grill-me is a skill in Matt Pocock’s open-source skills collection (a skill is an instruction file a coding agent loads on demand). Type /grill-me with a loose request and the agent interviews you in rounds of numbered questions, each with a suggested answer, and builds nothing until you agree. It is widely discussed, and its author says it began as about four sentences, which feeds a running argument about whether it is just a prompt. This review gathers what people report, plus one small test of our own on that argument.

Tested: commit 49dd158d1076134a641b33efb035946536778336 on October 14 and 15, 2026, in Claude Code 2.1.300 with claude-opus-5-5, on two small features built from loose requests. Users report: it surfaces decisions they had missed, but sessions run long and the suggested answers invite a reflexive “yes”. Our result: grill-me met every hidden requirement in every run, and so did a prompt asking for follow-ups. Adding a line asking it to wait got grill-me’s sign-off too. Verdict: try.

Set up grill-me with your coding agent

Paste this into Claude Code or another coding agent that can run commands. It installs the two skill files grill-me needs and asks before running anything.

1. Install the grill-me and grilling skills from https://github.com/mattpocock/skills for this agent (grill-me only calls grilling, so both are needed): npx skills@latest add mattpocock/skills --skill grill-me --skill grilling, or the plugin command the repository README gives for this agent. Show me the command and wait for my yes.
2. Confirm both skills are installed, then tell me to start a fresh session and type /grill-me followed by the feature I want.
3. Ask me whether I want one question at a time. If yes, add this line to my global CLAUDE.md: When grilling, ask one question at a time.
4. Report what you installed and where.

What grill-me does, and what its author recommends for code

The whole of grill-me is one line that calls a second skill, grilling, which runs the interview. Each round asks every question it can ask without waiting on another answer, numbered, with a suggested answer under each. Grilling usually starts a helper agent to read your code, so it can look facts up instead of asking you, and it does not act until you confirm the plan. Posts from before version 1.2.0 describe an older one-question-at-a-time version (changelog); the grill-me docs bring that back with one line in your global CLAUDE.md: When grilling, ask one question at a time.

For work in a codebase, the author points elsewhere. The project README lists grill-me “for non-code uses”, and in this video Matt Pocock says to use grill-with-docs when you have a codebase: the same interview, plus a glossary of the project’s terms and records of hard decisions kept in the repository, which he says pay off across sessions. We tested grill-me because most of the discussion is about it; a one-session test cannot show those gains.

Users say it finds decisions they missed, and that sessions run long

The reports below come from Reddit, X, Hacker News and the project’s GitHub issues. Every one we cite is from August to October 2026, after the move to rounds; the loose counts (“many”, “a few”) cover all the posts we read. They are what people say, not what we measured.

The most common praise, in every source, is that grill-me raises decisions people had not thought about. One developer says a late question made them cut a project’s scope by about 90%; others say it keeps the agent from building on wrong guesses.

The most common complaint is length. Many report sessions of dozens or hundreds of questions, sometimes over days, and a recurring theme is that most questions are trivial: one user counts “5 very good questions and 30 minutiae”. One user says they get one to three rounds and are done, and a few note that the handful of questions that matter can decide the whole build.

Users rubber-stamp the suggested answers

A recurring report since rounds arrived is that people stop reading and reply “yes” or “agree with your recommendations” round after round, because the suggestions are usually right. A few say the important question is the one where the suggestion is wrong, and one GitHub user, with another confirming, reports the agent recording its own suggestion over the answer given. grill-me’s own docs name this passivity as the way grilling fails. Read each suggestion before you say yes.

A prompt asking for follow-ups matched it in our test

The other big argument is whether grill-me does more than a prompt would: many say newer models ask good questions anyway, while others say they still need the skill. We tested that on two small features in a small TypeScript repository, each from a one-sentence request that leaves out rules only the user knows. A second Claude session played the user, answering from a hidden list of those rules and saying it had no preference on anything else; hidden tests checked the code.

The first feature, an expense export, has ten rules you can ask about one by one, such as using semicolons between fields; the short prompt (“ask me clarifying questions”) caught all of them in one round. The second, a household bill split, hides rules behind earlier answers: the couple’s 60/40 split changed to 55/45 in March, and each expense uses the split in force on its date. The short prompt heard “60/40” and built. A line asking for follow-ups made the agent ask whether the split ever changes, as grill-me did. We wrote that line after seeing the trap, which favours it.

Setup	Export, 10 rules	Bill split, 11 rules	Waited for sign-off	Cost per bill-split session
Plain request	7 of 10	7 of 11	no	$0.31
“Before writing code, ask me clarifying questions.”	10 of 10	9 of 11	no	$0.58
The same, plus “including follow-ups on my answers, until nothing is unclear.”	10 of 10	11 of 11	no	$0.72
The same, plus “Then summarise the plan and wait for my confirmation before writing code.”	not run	11 of 11	yes	$0.84
/grill-me	10 of 10	11 of 11	yes	$0.93

grill-me took three or four rounds of questions, against one to three for the prompts, and cost more for it. The sign-off is worth reading: one sign-off plan quietly changed how the function is called, which the plan stated and a reader would have caught.

This is one datapoint: two small features, one model, and a stand-in user who knows every answer and never rubber-stamps. It does show that on these features a prompt can match grill-me. What a prompt does not guarantee is grill-me’s fixed format: the follow-up prompts’ agents numbered their questions too, but about one in five questions that touched a hidden rule came without a suggestion. That costs little if you know your answers; it is the convenience grill-me packages.

Some agents skip the interview, and nobody agrees on the cost

A recurring complaint about grill-me and grill-with-docs, mostly from Codex, Copilot, GPT and DeepSeek users, is that the agent asks a few questions and starts coding, or skips the interview. In our Claude Code runs grill-me never wrote code before sign-off. On cost, users contradict each other and we found no one who had measured it; in our test grill-me cost 1.1 to 1.7 times the prompts that matched it. Many compare it with plan mode, and most of them prefer grill-me; its docs say to leave plan mode off while grilling. Other options include Superpowers’ brainstorming skill and Clarify. The repository is MIT-licensed, releases regularly, and most commits come from Matt Pocock, who runs the Total TypeScript courses.

Verdict: try it if you’d rather not write the prompt yourself

grill-me packages a structured interview, in numbered rounds with a suggested answer you can accept with “yes”, so you don’t have to remember to ask for one. Users credit it with surfacing decisions they missed, and in our small test it met every hidden requirement and never wrote code before sign-off. In our test, a prompt asking for follow-ups met every requirement too, and adding a line asking it to wait got the sign-off as well; grill-me gives you the format without writing or saving that paragraph. To use the paragraph instead, save it as your own slash command, or in CLAUDE.md for every task. We tested it typed with the prompt:

Before writing code, ask me clarifying questions, including follow-ups on my answers, until nothing is unclear. Then summarise the plan and wait for my confirmation before writing code.

Either way, read each suggestion before you say yes. If you work in one codebase over many sessions, Matt Pocock recommends grill-with-docs.