If you have spent more than a few hours working with large language models, you already know the feeling. You tweak a prompt, run it again, and the output looks different. Sometimes better. Sometimes …