Grill Me
Matt Pocock's /grill-me agent skill: the AI interviews you in rounds about a loose idea until every design decision is resolved
Overview
Grill Me is a user-invoked agent skill from Matt Pocock's open-source mattpocock/skills collection. You type /grill-me with a loose idea, and the agent flips the usual flow: instead of writing a plan, it interviews you in rounds until every branch of the design is resolved. Each round covers the "frontier", the questions whose prerequisites you have already answered. It is stateless (it writes no files), works on any subject, not just code, and needs no repo or setup. It is part of a collection with strong, fast-growing adoption (see badge for the repo's count) that installs into Claude Code, Codex, GitHub Copilot, Gemini CLI, Cursor and more.
The Verdict
Who Should Use Grill Me?
Best For
- Developers who keep getting AI plans that miss what they actually meant
- Turning a vague feature idea into decisions before any code is written
- Non-code decisions too (product, writing, process), since it works on any subject
- People who will push back and say "I don't know" instead of agreeing to everything
Not Ideal For
- Very large scopes, which can produce hundreds of questions (split the work first, or use Wayfinder)
- Questions about look and feel that need a prototype to react to
- Anyone wanting a saved spec or decision log from the session (use grill-with-docs or to-spec)
- Quick fixes where a long interview costs more than it saves
What's Great
- Reverses the interview direction, which targets the main failure mode of AI coding: misalignment between you and the agent
- Rounds build on your answers, so you surface decisions you had not noticed you were making
- Stateless and setup-free: no repo, config or files needed
- Free and MIT-licensed, with installs for Claude Code, Codex, Copilot, Gemini CLI and, via
npx skills, Cursor, OpenCode, Windsurf, Amp and others - Hands off cleanly to the same collection's to-spec and prototype skills
Watch Out For
- Only as good as the model: the docs advise your best model for grilling, and results vary between models
- Long sessions can drift into the "dumb zone" where a full context window lowers question quality
- Passive answers ("agreed" to everything) produce a plan you did not really decide
- Leaves nothing on disk; starting a fresh session before writing the spec loses the work
- Must be invoked by hand, and plan mode should be off or the agent jumps to a plan
Pricing
View all features & details
How a Session Works
- Start a fresh conversation and type
/grill-mewith your idea - The agent asks questions in rounds, one frontier at a time
- You own the scope: push back, flag drift, say "I don't know"
- It ends when the frontier is empty and no assumptions are left unstated
Signs It Is Working
- You disagree with it at least once
- Questions build on your earlier answers
- You reach a decision you had not noticed you were making
- You can defend each choice to someone who was not there
Related Skills in the Collection
- grilling: the underlying interview skill both entry points use
- grill-with-docs: stateful version that reads a codebase and records a glossary and ADRs
- wayfinder: maps efforts too big for one session and runs grilling inside them
- to-spec and prototype: write the spec, or build a throwaway version for questions you cannot answer in words
Supported Agents
- Claude Code (official plugin marketplace)
- Codex and GitHub Copilot (plugin marketplace)
- Gemini CLI
- Cursor, OpenCode, Devin, Windsurf, Amp, pi (via
npx skills)