Good idea! We can have some sort of standard grammar that we use to prompt the LLM such that it deterministically gives us the result we ask for. We then constrain all prompts to match that grammar. Some sort of language describing programs.