grillme: The Minimal Skill That Forces AI Agents to Clarify Requirements Before Coding

grillme's strength lies in attacking, with extreme minimalism, the most fatal weakness of the AI Agent era.

It is not a large framework project but a Skill of only a few lines (mainly from former Vercel engineer Matt Pocock). The core idea is simple: before writing any code, it pins you to your chair and relentlessly interrogates you, question by question, along every branch of a design decision tree, until you and the AI reach genuine shared understanding.

Why is this so effective?

The typical death of an ordinary AI coding agent goes like this:

You say "I want to add a comments feature," and it immediately starts writing React components + API + database schema... When you look at the result: "That's not it — I wanted one with moderation, Markdown support, @-mentions... and it must work with our existing user system." Then rework, changed requirements, more rework.

grillme does the exact opposite: it doesn't start building; it first becomes an extremely demanding foreman.

It asks questions one by one:

• Who is this comments feature for? Do anonymous users need support?
• Is moderation automatic or manual? How are edge cases handled?
• Should Markdown be supported? Should @-mentions trigger real-time notifications?
• How does this integrate with the existing permission system? Any performance considerations?
• If we later add "nested replies," should the current model reserve extension points?

For each question, it also gives its own recommended answer, then waits for you to confirm or correct it. Whatever it can find in the codebase, it looks up itself instead of asking you repeatedly. Only after all hidden decision branches are clarified and both sides say "yes, we're thinking the same thing" does it let you start work.

This essentially **turns Fred Brooks' decision-tree thinking from *The Design of Design* into executable agent behavior.

What makes it truly powerful

1. Treating the root cause: The biggest waste in the AI era isn't "slow coding" but "starting work before human and model are aligned." grillme blocks that waste at the source.

2. Minimalism equals power: The prompt is absurdly short, yet it changes the paradigm of the whole workflow — a classic demonstration of "less is more."

3. Naturally high-quality output: Organized afterward, the entire interrogation transcript becomes a solid PRD + implementation plan. The history itself becomes documentation.

4. Especially friendly to heavy agent users: If you're working on KLIP, agent flows, context engineering, or self-evolving systems, you often face "vague but high-complexity" tasks. Using grillme to map out the decision tree first, then letting the agent execute, saves enormous hidden costs.

It's not about "being dominated by AI" — it forces you and the AI to get things clear before doing the work. Real engineering masters are never the fastest to start; they're the ones who figure out "what exactly are we building" first.

In short: grillme is great because it re-implements the old principle of "think first" in the most efficient way of the agent era.**