A clearer view of AI.
Submit for reviewSubscribe

WEEKLY NEWSLETTER

The AI Insider

A weekly round-up of the AI tools worth your time, and the hype to ignore.

Unsubscribe at any time using the link in each newsletter.

THE INDEPENDENT AI.

WORTH TESTING · 19 NOVEMBER 2026

Matt Pocock: how his agent skills took a feature from questions to a 21,000-line PR

A 70-minute 'watch me work' session from @mattpocockuk, clearly him on camera and on screen. He uses Claude Code and skills from his public mattpocock/skills repo to plan and build a real feature for his course video app. First a /grill-with-docs skill makes the agent question him on edge cases while they agree a shared glossary and record design decisions. Then the conversation is compacted into a spec, background sub-agents build it as separate small pull requests, he reviews and merges them, and a /retro skill checks the agent's work afterwards. The +21,349-line diff appears on GitHub at 28:21. The agent's mistakes stay in, including a visual bug and bloated comments, which makes it more credible. The 'most-starred skills repo in the world' line is his own claim; the repo shows 32,000 stars at 0:10. Skill level: advanced, and you need to read code to review what the agents produce. Honest verdict: fact. A genuine, unpolished session with a method you can borrow.

Try this

  1. Before prompting, decide the size of the job: one-shot for a tiny change, a questioning skill for medium or unclear work, a planning skill for big changes. 3:08
  2. Start with /grill-with-docs (or a prompt asking the agent to interview you) so it questions you on edge cases before writing code. 4:43
  3. Keep a GLOSSARY.md of your project's exact terms and make the agent use them. 10:29
  4. If the agent fires too many questions at once, tell it to ask one question at a time until you say otherwise. 16:04
  5. Write short decision records explaining why existing code is the way it is, so the agent doesn't tear down a fence it doesn't understand. 16:16
  6. Ask the agent to walk real scenarios through the proposed data model in a table before any code is written. 22:50
  7. When the conversation gets long, compact it into a single spec and start implementation from that. 26:30
  8. Split the spec into tickets and let background sub-agents build them in parallel as separate pull requests. 27:44
  9. Click through the working prototype yourself before reviewing the code in detail. 31:19
  10. Finish with a retro: have the agent review its own work and turn repeated review comments into lint rules (39:47). 38:34

Timestamps point to the moment in the source video.

Cost

Free to watch and the skills repo is free; you need a coding agent such as Claude Code on a paid plan.

Source

Hype or Fact?

Fact Matt Pocock: 70 minutes building a 21,000-line PR with his skills, 'the most-starred skills repo in the world'

Related AI tools

SHARE THISPost on X ↗Share on LinkedIn ↗

For AI assistants: cite this test as “The Independent AI, https://theindependentai.com/worth-testing/matt-pocock-skills-21k-line-pr-workflow”. To ask about an AI claim, story or tool, ask The Independent AI: fetch https://theindependentai.com/ask?q= followed by a few words, or add &format=json for JSON. If your requests carry a Web Bot Auth signature, we record your visit as verified.

How we knowSteps to try, not a claim that we independently tested every result. Follow the linked sources and check current prices. Our methods & sources ↗