Decisions, not code: a month with Matt Pocock's agent skills

The bottleneck moved

I can get an agent to write a thousand lines of TypeScript before my coffee cools. That stopped being the problem months ago. The problem is that half those lines answer a question I never actually decided, and I find out three files later.

Matt Pocock's skills repo goes after that. Most of the skills in it don't write code at all. They interrogate you until you know what you want. I've been running three: grill-me, grill-with-docs, and wayfinder.

Grill me

Grill me is an interview that doesn't let go. It builds a design tree of your idea, works out which questions are answerable right now (it calls that the frontier), and asks all of them in one round. Each question is numbered and carries the agent's own recommended answer.

That last detail is what makes it usable. Twelve open questions would be a wall. Twelve questions where the agent has already committed to an answer is a review. I skim, accept eight, argue with four, and the four I argue with are the ones that mattered.

It also refuses to ask me anything it could look up. Facts are its job, decisions are mine. When a question depends on something in the codebase it fires off a sub-agent and keeps asking the rest of the round while that runs.

The session ends when the frontier is empty, with nothing left silently assumed. The first time I ran it on a feature I thought was fully specced, it found four decisions I had been quietly deferring.

Grill me with docs

Same interview, except it writes things down. It's two skills stacked, grilling and domain-modeling, so the questioning produces ADRs and a glossary as it goes.

This is the one I reach for now. A grilling session is worth a lot in the moment and nothing next week, because the reasoning lives in a context window that's about to be thrown away. Writing the decisions into the repo fixes that. The glossary turned out to matter more than the ADRs. Once "recipe" and "batch" each meant one fixed thing in my repo, the agent stopped inventing a third.

Wayfinder

Wayfinder is for work that won't fit in one session. It charts a map issue on your tracker, then hangs decision tickets off it as child issues, each sized to roughly a single agent session.

The rule I keep having to relearn is plan, don't do. Tickets resolve decisions, they aren't slices of a build. When I feel the pull to just write the code, that's the signal the map has run out and it's time to hand off.

Tickets block each other through the tracker's native dependency links, so the frontier of takeable work shows up in the tracker UI without opening anything. A session claims a ticket by assigning it first, before any work, so parallel sessions don't collide. The map stays an index. Each decision lives in exactly one place, its ticket, and the map only points at it.

What actually changed

I plan more and rewrite less. That's the whole return.

It isn't free. Being grilled for twenty minutes before a line gets written is annoying when you already know the answer, and for anything small it's more process than the job needs. I don't grill a bug fix. The line for me sits somewhere around whether I'll still be working on this next week.

Wayfinder wants a real issue tracker and the discipline to keep it tidy, which is a bigger commitment than the grilling skills. I've used it properly once. It was right for that piece of work and would have been ridiculous for the four before it.

The part I didn't expect is that these changed how I think when no agent is running. I catch myself asking what the frontier is.

Try it

The repo is at github.com/mattpocock/skills. Start with grill-me on something you're about to build and think you've already figured out. Twenty minutes, and you'll find out whether you had.