CyrilXBT|Nov 22, 2026 09:16
IF YOU'RE STILL BABYSITTING YOUR AI AGENT, THE PROBLEM ISN'T THE AGENT. IT'S YOU.
Matt Pocock just shipped a 21,000 line feature without hovering over every step.
Then he wrote down exactly how in a 10 page PDF.
His skills repo reportedly has more stars than React.
Here's the formula:
Feature = Codebase + Grill + Spec + Tickets + AFK Loop + Retro
Most people hand the agent the "what" and then micromanage every step of the "how."
He does the opposite.
The 10 layers:
0. Codebase: deep modules and fast feedback. Bad codebases make bad agents.
1. Setup: tell the skills how your repo works, once.
2. Grill: the agent interviews you until every question is answered. Give it the why.
3. Glossary: one word per idea. The reasons go in decision records.
4. Prototype: 3 throwaway versions. Keep the decision, not the code.
5. Spec to tickets: small vertical slices, each one demoable.
6. AFK loop: the agent grabs the next ticket, writes tests first, commits, repeats overnight.
7. Review: a fresh reviewer checks your standards before you even look.
8. Stack and QA: humans only review the decisions you can't undo.
9. Retro: every review comment becomes a lint rule.
The result:
47 user stories.
9 parallel tickets.
About 21,000 lines.
All 12 review comments turned into permanent rules.
Here's my take.
Most developers complaining that AI writes bad code never gave it a good codebase, a clear spec or a single rule.
Then they blame the model.
The agent was never the bottleneck.
The process was.
Share To
Timeline
HotFlash
APP
X
Telegram
CopyLink