if you are using contemporary LLMs, you want prompts that amount to linguistic one-way functions. they should generate output that would otherwise be difficult or annoying or time-consuming to produce, but is trivial to verify.