The grill me skill turns your AI agent into the colleague who asks the awkward questions before you start building. You describe a plan, and it interviews you about it, a round of questions at a time, until every decision has an answer. It is one of the most installed skills on skills.sh, and the idea is simple enough to explain in a sentence: settle the plan before you write the code.
What it does
The agent treats your plan as a tree of decisions: some can be answered now, and some depend on others. Each round it asks everything that can be answered now, numbered, with its own suggested answer under each question. You reply, the tree changes, and it asks the next round.
Two rules keep it honest. Facts are the agent's job: if a question needs something it can look up in your code or tools, it looks it up rather than asking you. Decisions are yours: it puts each one to you and waits. From the skill:
Finding facts is your job, never the user's. [...] The decisions are the user's: put each to them and wait.
From SKILL.md by Matt Pocock, MIT.
It's finished when nothing is left undecided, and it won't act on the plan until you confirm you both understand it the same way.
An example
You say: "Grill me: I want to add comments to my blog." The first round might ask whether readers need an account to comment, whether comments go live immediately or wait for approval, and where they're stored, each with a suggested answer (no account, approval first, your existing database). Once you answer "approval first", the next round opens questions that only exist because of that answer: who approves, and how they find out there's something waiting.
When to use it, and when not to
Use it before any change with more than one moving part: a new feature, a data model, a project you're about to start, a pull request you're not sure about. It works for non-code plans too.
Skip it for small, clear changes. Being grilled over a typo fix wastes everyone's time.
What we checked
We read every file in the grilling folder at the commit linked above:
SKILL.md and agents/openai.yaml (a display name for Codex). We also read the
grill-me shortcut next to it. There are no scripts, no network access and no
allowed-tools. The skill can start a sub-agent to look things up in your
project; that sub-agent has the same permissions your agent already has. The
repo is MIT licensed. Our full checklist is in the
guide to checking a skill before you install it.
When the plan is settled and built, Matt Pocock's code review skill checks the result against what you asked for.