Build a prompt out of the parts that stop a model being confidently wrong, and keep the ones that work.