Jason Zhou of AI Builder Club has pulled Andrej Karpathy’s ideas, scattered across talks and posts, into a single playbook for building with AI. I read it as someone who puts agents to work in companies that are nothing like a Silicon Valley startup. Three things I take straight away. Two I’d handle with care.
First, a story about a yellow button
This week I rebuilt my website together with an AI agent. There’s a small 3D factory on it, and a man standing at the conveyor belt who presses a big yellow button.
The agent built him fast. Then I spent half an evening sending short notes: don’t spin on the spot, don’t stand on the cable, the palm has to lie flat on the button. Each fix took it a few minutes. It didn’t come up with a single one on its own.
That is the difference Karpathy talks about, in miniature. You can throw in a prompt and hope. He once called that “vibe coding”. Or you can describe what you want, look at the result and say what to fix. The playbook calls the second way agentic engineering.
What I take straight away
Rules written down once. Karpathy keeps a short file of rules for his agent. In a company it works like a job instruction. Mine has lines such as: prices in dollars on the English site and in złoty on the Polish one, and we ask people to get in touch by email, never by phone. Without that file I would repeat it in every conversation, and every so often I’d forget.
A person reads the changes instead of writing everything. The agent takes the first pass. I look at what it changed and decide whether it goes in. It’s a different job from writing from scratch, and it has to be learned. The hard part is spotting a mistake in something that looks fine at first glance.
Notes that grow. Karpathy described a setup where the model keeps organised notes and adds new findings to them, instead of searching the documents from the beginning every time. I do something similar: the agent writes what we agreed into files, and the next agent starts by reading them. A month in, I don’t explain anything twice.
Where I’d be careful
Several agents at once. According to the playbook, Karpathy runs several sessions in parallel, one per feature, like a small team. For a programmer with his experience that makes sense. In an ordinary company the bottleneck quickly becomes the person who has to check it all. So I start with one agent and one process.
“The context is the program.” The line captures how working with models feels today: what the model gets to read decides what comes out. But in a company half of the context lives in people’s heads. Before an agent can take anything over, someone has to write it down, and that is usually the longest part of the project.
What to do on Monday
- Pick one repetitive task that nobody on the team enjoys.
- Write one page of rules a good employee follows when doing it.
- Decide who checks the agent’s results, and for how many weeks. I wrote more about that in Your AI agent is on probation.
Zhou’s playbook is worth reading in full, if only to see how these ideas fit together. I’m staying with the version for an ordinary company: one task, one page of rules, and one person who keeps an eye on the agent.
Want to find out which task in your company is a good place to start? Book a 20-minute online meeting.
