@denisfi here is an example how to use (local) LLMs to do chunks of tasks and collect the results in JSON structures to (re-)use them later. Maybe you can adapt that approach.
The functions depend very much on the sort of LLM and the prompt. One idea can be to let AI write the prompt.