First thing I noticed: every one of them writes prompts that assume the model will fail. Not optimistic prompts. Adversarial ones. They're not trying to get a good answer. They're trying to catch where the model breaks. That changes everything about how you write prompts.

→ View original post on X — @godofprompt