How do you test that a prompt still works after a model update?

I have about thirty prompts running in production features. Every time a model version ships I find out something drifted by a customer complaining, which is not a great detection system. What I am looking for is a practical routine, not theory: - How do you decide which prompts are worth regression testing? - What do you assert on, given the output is not deterministic? - Do you keep golden outputs, or check properties instead? If you have a prompt that generates test cases for another prompt, that would solve half of this for me. Interested in what people actually do rather than what sounds rigorous.