← writing

Most of my rules are prose

1 min read

I went through Lauren Tan’s talk All Code Is Slop (Except Mine), about getting reliable code out of AI agents. A lot of it felt familiar. I got to some of the same habits on my own, mostly by correcting the same agent mistakes again and again: rules in CLAUDE.md, a lessons file in most projects, reviewer agents that see only the diff and the spec, and vir, which turns my sessions into notes the agent reads next time.

Her talk also put those in their place. She describes a pyramid: types, lint rules and tests at the bottom, then verification, and prose rules at the very top, because they are the least reliable. Most of what I built sits in that top layer.

Some of it is lower. This site’s CI fails on any extra JavaScript or a page over budget, and Fulcrum has 199 tests and an attack suite. But vir still writes lessons down as prose. In her pstack skills, a mistake that happens twice gets fixed in the architecture or the types first, then with a lint rule whose error names the fix, then a test. Docs come last, only for judgment calls.

That’s the part I want to borrow.