Prompting as pattern design (why formatting beats adjectives)

Before models were trained to behave like assistants, “prompting” wasn’t really instruction-following. It was closer to genre emulation. You’d get reliability not by saying “be helpful” or “be professional,” but by giving the model a pattern it had seen a thousand times—a demo script, a Q&A template, a document header, a list with 1. already started.

That’s why separators, schemas, and rigid formatting mattered so much. They didn’t just make the prompt readable to humans; they put the model into a familiar groove. In that era, the winning move wasn’t better adjectives. It was better structure.

The core lesson that survives today: when you want consistent behavior, change the interface before you add more instructions.