Prompting rules that stopped working
Everyone publishes what the models can do. Nobody publishes what they stopped needing you to say. These are instructions that were good advice on an older model and now actively hurt — the ones that make your agent slower, more expensive, or quietly worse.
Every item is dated. When one turns out to be wrong I retract it here rather than deleting it, because a list that only ever grows is a list nobody is checking.
How long each of these lasts
- law — a principle about how models change. Survives releases.
- api-fact — falsifiable in a single API call. Either true or it isn't.
- market — true until a vendor ships something. These are the ones that go actively false, and the ones I retract most.
Get the rest
This list ships inside the prompt pack, along with 22 prompts grouped by job, a model cheat sheet, and a CLAUDE.md template. I send an update when the list changes — including what I've retracted.