It gets more interesting when you get to benchmarking your prompts for accuracy. If you don't have an evaluation set you are flying blind. Any model update or small fix could break edge cases while you don't know.