A light process for working with AI coding assistants that makes the output reviewable, resumable, and trustworthy. The code isn't the interesting part. The process is.
AI-assisted coding works fine until it doesn't. Pasting a vague prompt into a chat window builds a toy. Building something you'll still understand six months from now takes a bit of structure. The process below takes that structure and makes three things true at once.
A fresh LLM session, another vibe-coder, or an engineer who's never opened the repo can read a handful of commits and understand exactly what happened — what was asked for, what was tested, what shipped. "Anyone" really does mean anyone.
Any piece of work can be picked up months later from a single short document — by me, by a collaborator, or by a new LLM session that has never touched the repo. The "why" isn't lost to memory, and the original intent is still legible in the code.
The assistant can't quietly break boundaries the tests enforce. So I don't have to hand-check every diff — just the parts where judgment mattered.
None of this is clever on its own. None of it requires special tooling. What it does together is make the output of an LLM-assisted session something you can hand to anyone — a collaborator, a fresh LLM, or yourself six months later — and have a good time reading it.