← Back to Journal

Matt Pocock's AI Coding Workflow

OCTOBER 19, 2026 · Engineering

I watched Matt Pocock's video on his AI coding workflow, which covers version 1.3 of his open-source agent skills repo. I'm starting a role as an AI Agent & Workflow Developer, so I wanted to see how someone who uses coding agents every day runs them.

The problem

Ask an AI agent for a "pause subscription" button and it writes the code fast, but it guesses the business logic. What happens if someone pauses after their coffee bag is already packed? Does billing stop today or at the end of the month? The agent picks an answer and keeps going.

Grill me

The first step is a skill called Grill Me. Before writing any code, the agent interviews me. It treats the feature as a decision tree: the feature is the root, and each open question (billing timing, already-packed orders, how long a pause can last) is a branch. It only asks about the decisions whose parent choices are already settled, which he calls the frontier, and it won't write code until the frontier is empty.

The questions come in numbered rounds, and each one includes the agent's recommended answer, so I can reply "yes, yes, no, end of the month" instead of writing paragraphs. For things that can't be answered in words, like what a screen should look like, the agent builds a few throwaway prototypes in a fresh session. I click through them, pick one, and the answer goes back into the interview.

A shared glossary

A variant, Grill with Docs, writes the project's terms into a glossary in the repo as the interview goes. That stops the agent from calling the same thing a "pause," a "freeze," and a "hold." If a message is confusing, typing "wait what" makes it restate its last point in plain English using the glossary's terms.

Small tickets, fresh sessions

Once the plan is settled, it gets cut into small tickets, and each ticket runs in its own fresh chat from start to finish. Handing the agent a huge feature all at once is where its output gets worse.

Review by someone else

The agent that wrote the code never reviews it, since it tends to think its own work is fine. Two fresh reviewer agents do that instead: one checks that the code matches what I asked for, and the other looks for code smells and standards violations.

Retro

When the feature is done, I type retro. The agent looks back at where it struggled, like spending ten minutes hunting for a file, and suggests a rule update or a lint check so it doesn't repeat the mistake.

What I took from it

His framing is that I'm the general and the AI is a fast sergeant: it does the typing, and I make the strategic calls. The interview step is the part I'd copy first, because it surfaces decisions I'd otherwise only find after the code was written.

I haven't tried the skills yet. I want to run Grill Me on my next feature, before I let an agent write anything.

Engineering