How do you actually know if a prompt change made things better or worse?

I went through Anthropic's Academy a while back and it changed how I think about prompts. The thing that stuck with me: a prompt can look fine on one or two tries and still fall apart the moment the input changes, and most of us never actually check that.

I tried testing one of my own prompts properly, against a real set of cases instead of the two or three I'd normally glance at. It scored 4.9 out of 10, zero percent pass rate. Changed one thing (being explicit about the output format) and it jumped to mostly 8 out of 10.

That gap is what got me building Prompt Eval, a way to run a prompt against a test set and get an actual score instead of a gut feeling.

Curious how others handle this. Do you test prompt changes systematically, or is it mostly "run it a couple times and see if it looks right"?