A 21,000-Line PR: Matt Pocock Spends 70 Minutes Unpacking the Agent Workflow
Matt Pocock released a 70-minute hands-on video, building a 21,000-line PR live, with a focus on DDD, /retro, and human review. The comments immediately seized on the awkward question: with a PR this large, who is actually reading it?
Matt Pocock posted a 70-minute video in which he builds a 21,000-line PR live. The original post is here: https://x.com/mattpocockuk/status/2123098316867051991
In the tweet, he says it has been almost 6 months since his last watch me work deep dive. During that time, his skills repo became the most starred skills repo in the world. The video focuses on three things: using /grill-with-docs for DDD; using /retro to improve the codebase; and how to conduct effective human review.
In a reply to the tweet, he included a free email course, AI Skills for Real Engineers, with 7 lessons; the page shows 135,400+ developers subscribed. The course covers everything from using /grill-with-docs to clarify requirements, to /prototype and /handoff to validate uncertain ideas, to /to-spec and /to-tickets to break work into tasks, running an AFK agent with Sandcastle, and finally wrapping up with /code-review. Link: https://www.aihero.dev/skills/subscribe
The course page makes Matt's identity clear: creator of AI Hero and Total TypeScript, with a quote from Guillermo Rauch. The page also lists tutorials on Vercel AI SDK, MCP, and more. For people unfamiliar with him, this information is enough.
The comments are more direct than the video title. Khoa says that with 21,000 lines, the review portion is no small burden. Arjun asks, did you really read all 21k? Kirill says he would have Claude watch the video and summarize it. Freddy asks how to track what the agent actually changed. Justice U's reply is the most specific: 21,000 lines means no one will read the entire PR; human review is really about choosing what to read—starting with tests, module boundaries, and changes related to payments or authentication. Sebastian says he used Matt's skill stack for three months and was later inspired to build his own stack.
Some people also poured cold water on it. op says that with the latest Claude models, you don't actually need skillmaxx. That judgment is debatable, but it points to a reality: the stronger the models get, the more the value of process has to be proven. What the 70-minute video lays out is a reusable workflow, and how review can stay under control in an enormous diff. With a 21,000-line PR sitting there, line-by-line reading is no longer realistic; what the process needs to do is make sampling justified and let logs, tests, and boundaries expose risk.
For developers, the video is worth watching for the concrete operations: how DDD is advanced by /grill-with-docs, how /retro feeds back into the codebase, and how human review shifts from 'reading through it' to selective inspection. For general readers, at least one thing becomes clear: once agents can write code faster, the bottleneck moves to review and boundary design.
发布时间: 2026-11-19 01:26