// the field manual
How to build with AI that survives production
Seven lessons on building real software with AI, in order. Each one takes a practice the frontier teams publish and shows what it looks like in an actual build. They stack: every lesson leans on the one before it. Start at the top.
- 01Spec before you generateDecide what can't drift before the AI writes a line. The difference between shipping and rebuilding is a spec, not a better prompt.
- 02Start simpleThe complexity you add is a loan. Only take it when the simple version has actually failed you.
- 03Verify, don't trustGeneration is cheap now. The review is the actual work. A green build proves almost nothing.
- 04Make the bad state impossibleIf it must never happen, don't ask the model to be careful. Build it so it can't.
- 05Reversible by default, gate the irreversibleLet the agent move fast on anything you can undo. Put a human on the one button you can't take back.
- 06When it breaks, triage before you rebuildThe rewrite feels faster and almost never is. Diagnose the one line before you throw away the whole thing.
- 07Externalize your memory// unlock with emailThe context window fills and forgets. The move is to keep your working state somewhere it can't.