How I A/B test LLM prompts without fooling myself

A while back I was building a support assistant and hit a simple-sounding question: is this new version of the prompt actually better than the old one? So I did the obvious thing. I wrote thirty test ...