grill-me: The Claude Code Skill That Interviews You Before Writing Any Code

Tell an AI coding agent to “add a login page” and it will build one. It just might not be the one in your head.

That gap between what got typed and what got meant is where most wasted coding-agent sessions come from, and it’s exactly what Matt Pocock’s grill-me skill for Claude Code was built to close.

What is grill-me?

grill-me is a Claude Code skill that reverses the usual direction of a prompt. Instead of the developer issuing an instruction and the AI executing it, the AI asks questions first, one at a time, and won’t start writing code until the plan underneath the request is actually clear.

The name is literal. Running /grill-me puts the developer in the hot seat, not the AI. It’s built specifically for the moment before code gets written, not for reviewing code that already exists.

Why does grill-me ask questions instead of taking orders?

Prompt quality varies wildly depending on how much domain-specific detail, terminology, and constraints the developer actually spells out, and most people don’t spell out nearly enough on a first pass. A one-line instruction like “build a booking system” leaves out who’s using it, what scale it needs to handle, and which edge cases matter, so the AI fills those gaps with guesses.

Sometimes those guesses land. When the task is simple enough or well-represented in training data, a vague prompt still produces something usable. When it isn’t, the result is code that technically runs but solves the wrong problem, and the fix costs more time and more tokens than answering a few questions upfront would have.

Who is grill-me built for?

grill-me fits people who want to nail down the system architecture themselves before an AI touches any code, and non-developers who understand what they want conceptually but haven’t translated it into a spec yet. Vibe coders planning a project before writing a single line benefit the most, since the interview format forces the kind of specificity that usually only comes from experience.

It’s a poor fit for anyone who wants to hand a task off entirely and check back later. The value of grill-me is the friction: skip the questions and you’re back to the same gap the tool exists to close.

How do you install and run grill-me?

Node.js needs to be installed first. From there, running npx skills@latest add mattpocock/skills in a terminal installs the skill alongside Pocock’s other Claude Code skills. Once it’s installed, open a project directory, start Claude Code, and type /grill-me. The AI reviews the directory first, then starts asking questions.

What does a real session actually look like?

The questions don’t arrive as a checklist. They chain: an answer about who the users are leads to a follow-up about expected scale, which leads to a question about what happens when that scale gets exceeded. A typical exchange might go something like this:

AI: What kind of users will this booking system serve? Dev: Small gyms, maybe 50 to 200 members each. AI: At that size, do double-bookings for the same time slot need to be prevented at the database level, or is manual conflict resolution acceptable?

That kind of question rarely gets asked in a single upfront prompt, and it’s the exact class of decision that’s expensive to unwind after the code already exists.

The friction is real too. Sessions with a lot of open questions turn into a long run of answer, hit enter, answer again, and that rhythm annoys anyone hoping to get straight to code. That’s the tradeoff: grill-me trades a slower start for fewer rewrites later, and it only pays off if you’re willing to sit through the slower start.

References

• GitHub: mattpocock/skills
• Node.js