Most people treat their AI prompt like a sticky note. “Be helpful. Be concise. Don’t lie.” Then wonder why the output sounds like a LinkedIn post written by a committee.

I spent an embarrassing amount of time doing exactly that.

A prompt isn’t a note. It’s architecture. And if you don’t design it like one, you get what you deserve.

So I started treating it like a design system. Define the tokens first - tone, register, epistemics. Then the components: how it handles corrections, when it asks questions. Then the constraints. The order matters. Components built on undefined tokens are just vibes with extra steps.

Tone is where most people underinvest. “Be casual” means nothing. Be casual like who, exactly? You have to get specific enough that you’re basically describing a person.

Epistemics was the part that actually hurt. Default AI behavior is to agree, hedge, then agree again in different words. Explicit instruction broke that. Conflict resolution order finished it - when brevity and correctness clash, correctness wins. No ambiguity, no hoping.

Does it work perfectly? No. It works well enough that I catch the failures. That’s the actual goal.