What Should a Prompt Regression Test Protect?

A prompt can change without breaking its output format—and still change what the workflow does. A response may remain valid JSON while assigning the next step to the wrong owner, skipping a required approval, or sounding certain where the evidence is incomplete.

When you test a prompt revision, which behavior should block release? Do you prioritize schema validity, correct routing and permissions, how uncertainty is handled, or examples of past failures? And when those checks disagree, who owns the decision to ship or revise?

I’m interested in how teams keep regression tests useful without turning every wording change into a full review of everything.