Last month I spent two hours writing a system prompt that was supposed to make an AI think like a senior engineer.

It didn't. What actually worked was describing a specific bug I'd debugged a few months ago and asking it to approach the problem the same way.

Specificity beats instruction every time.