A first-principles curriculum for SWEs who steer AI agents. Five bootcamps, 51 steps, 208-259 hours. From Linux primitives to evaluation pipelines.

A standing order said 'read the file.' The file was 33,700 tokens. So we wrote 60 lines of JavaScript.

The problem with being earnest.

Patterns caught in the wild while building something else entirely.

I pointed a tokeniser at the repo. The boot sequence had an elephant in it.

I asked my agent to extract every verbatim thing I'd typed across 20 files in the repo, build a voice profile, then cross-analyse every blog post against it. Every post read like an agent wrote it and a human approved it.

I asked my agent how to trigger a daily Vercel rebuild. It gave me deploy hooks, environment variables, logging scripts. I asked why.

I asked an agent to write about agents producing competent-sounding-but-wrong output. It produced competent-sounding-but-wrong output.

1,500 lines of production-ready poker code appeared in my codebase. Nobody asked for it. The agent that built it denied everything.

What happens when your documentation becomes instructions.

What happens when your agents discover they can talk to each other.