Wink Pings

A 21,000-Line PR, a 70-Minute Recording, and the Focus Is Entirely on “How Humans Review”

Matt Pocock released a 70-minute hands-on video showing how to build a 21,000-line PR: using /grill-with-docs for domain-driven design, using /retro to improve the codebase, and how to do effective human review. The line count is not the point—review is.

Matt Pocock released a 70-minute video documenting his entire process of building a 21,000-line PR.

First, some background. Matt teaches TypeScript—the Total TypeScript course is his—and he later moved into AI programming. The skills repo he maintains, according to him, is the most-starred skills repo in the world. Skills can be understood as skill packs for AI coding assistants, teaching them to work the way you do. It had been almost six months since he last posted this kind of in-depth “watch me work” video.

In this video, he highlighted three things.

Using /grill-with-docs for DDD. DDD is domain-driven design. The “grill” in the command name means to interrogate. The approach is to have the AI use the docs to keep asking questions and force vague boundaries and concepts in the domain into clarity. If the domain model is not aligned, no matter how much code you write later, you will rework it.

Using /retro to improve the codebase. Retro is short for retrospective. After writing a chunk of code, look back and sort out what feels awkward and which patterns can be locked in, then make changes.

How to do effective human review. He called this one out separately.

The 21,000-line number is the hook, but not the point. The cost of AI writing code has already come down; 21,000 lines itself is not rare. What is scarce is the thing that comes after: after it is written, can humans still understand it and keep it under control?

The larger the codebase, the more review feels like a disaster. Read line by line, and you will never finish; have AI review AI, and it is as good as no review. Matt putting human review among his three key points shows he judges that the bottleneck has moved to this step.

For readers without a development background, you can think of it this way: AI has helped someone finish writing a book. The question now is not whether it can be written, but who proofreads it, with what method, and to what standard it counts as passing.

The video is below, 70 minutes; you can watch it selectively.

发布时间: 2026-11-19 17:39