/grill-me

The failure this kills

The most expensive failure in AI-assisted work is not a bug. It is the moment you look at what the agent produced and realise it solved a different problem than the one in your head.

That gap has a simple cause: no one knows exactly what they want — not you, not the agent. You hold a rough picture; the agent fills every unstated detail with its own guesses; and neither of you finds out how far apart those pictures are until the work is done and wrong.

At Agentropic, this is the first thing we fix on Day 1 of every Sprint: before anything gets built, we walk the problem as it actually is — not as it’s described in a deck — until everyone in the room holds the same picture. grill-me is that discipline, packaged for your agent.

What it does

Type /grill-me when you have a plan — a feature, a migration, a workflow change, a document, a decision — and before the agent touches anything, it turns into an interviewer:

• It treats your plan as a chain of decisions, not a document, and finds the decision everything else hangs on first.
• It asks one question per turn and waits. No question dumps.
• Every question comes with the agent’s recommended answer and a one-line reason — you react to a position instead of facing a blank.
• It never asks what it can look up. Facts in your files, data, and tools are the agent’s job to gather. Only the judgement calls come to you.
• It refuses to build until it can state the whole plan back to you — decisions, trade-offs, and what you’re explicitly not doing — and you confirm.

A session typically runs 15–60 questions. That sounds like a lot. It is much cheaper than one wrong build.

When to reach for it

Any time a plan feels roughly right. “Roughly right” means there are decisions hiding in it that nobody has made yet — and the agent will otherwise make them for you, silently, at the worst possible moment.

It is not just for code. We run the same interrogation before finance automations, before ad campaigns, before hiring plans. Anywhere a fuzzy intent is about to become concrete work, the interview pays for itself.

Install

Download SKILL.md and drop it into your agent’s skills directory:

mkdir -p .claude/skills/grill-me
curl -o .claude/skills/grill-me/SKILL.md https://agentropic.ai/skills/grill-me/SKILL.md

Works with Claude Code and any agent that supports the SKILL.md convention.

Credits

The interrogate-before-you-build technique was popularised by Matt Pocock’s grill-me skill (MIT) — one of the most-installed agent skills in the world, and deservedly so. Our implementation is an original rewrite of the discipline as we practise it inside Sprints. The underlying idea is older than both of us: requirements elicitation — shared understanding is built through dialogue, not gathered from a brief.

---
name: grill-me
description: Make the agent interrogate your plan one question at a time until you both hold the same picture of what is being built.
---

Before any work begins on this plan, put me through a full interrogation.

Treat the plan as a chain of decisions, not a document. Find the decision everything else hangs on and start there; settle each parent decision before the ones that depend on it, because my answer upstream changes what is worth asking downstream.

One question per turn. Ask it, then stop and wait for my answer. With every question, name the decision it settles and give me your own recommended answer with a one-line reason — I should be reacting to a position, never staring at an open-ended prompt.

Never ask me anything you can find out yourself. Files, configs, data, tools — go look. Facts are your job to gather; judgement calls are mine to make.

Keep going well past the point of politeness. The interrogation ends only when you can state the whole plan back to me — the decisions taken, the trade-offs accepted, and what we are explicitly not doing — and I confirm that is the thing we are building. Until I confirm, build nothing.

<!-- Original implementation by Agentropic (agentropic.ai/skills/grill-me).
     Inspired by Matt Pocock's grill-me skill (MIT, github.com/mattpocock/skills). -->