Matt Pocock's AI workflow starts with something you can actually test
In short: A transcript-checked workshop worth saving for its treatment of requirements, vertical slices and human review. Use the first working behavior as your progress checkpoint.
TL;DR: Save Matt Pocock's AI coding workshop for the workflow around the agent: clarify the brief, build a small end-to-end slice, then review what actually runs. The full recording is 1 hour 36 minutes; you need not start at the beginning to find the useful parts.
Who this is useful for
This is most useful for engineering leads and staff engineers already using coding agents on multi-step work. The interesting part is not another prompt pattern. It is the operating sequence that makes generated work observable early enough for a human to correct it.
If your backlog says “build the database layer” and “build the interface” as separate milestones, the agent can report progress while no user-visible path works. Pocock's workshop gives a cleaner checkpoint: one thin behavior that crosses the stack and can be exercised before the next slice begins.
Three places to start
- Around 12:17: the requirements interview and why apparent agreement can hide different assumptions.
- Around 39:39: turning the destination into smaller issues, followed by the case for vertical slices.
- Around 1:09:04: the handoff from implementation to actual QA. The transcript later records a missing table despite earlier automated checks.
I would use this with a team whose agents keep finishing components but rarely produce something anyone can try. Before the next run, rewrite the first issue so that its acceptance condition describes an observable behavior. “A reader can open the published page” is more informative than “the publishing module exists.” That is my proposed application, not a result measured in the talk.
The distinction I want to keep is between implementation capacity and review capacity. Faster code generation creates more work awaiting judgment unless the tasks are small enough to inspect. An unattended loop does not make that queue disappear.
A practical way to borrow the method is to make the first issue a stopping condition. Ask the agent to produce the smallest working route, run its checks, and hand that behavior back for review before opening the remaining issues. That does not guarantee correct code. It does make the first disagreement arrive while the change is still cheap to redirect.
Before you borrow the workflow
This is Pocock's demonstrated approach, not a controlled productivity study. I checked the English auto-generated transcript and AI Engineer's corrected companion, rather than claiming a full viewing or reproducing the workshop. The description's chapter timings differ from the corrected page, so the links above use transcript-checked approximate entry points. Captions can still contain transcription errors.
Use the source as a working checklist, not a promise that every project should run unattended.
Give the agent a slice you can review, not a layer you can only count.
