Build a mail parsing pipeline. Return JSON {"pipeline": [list of 2+ steps], "why": "short"}.
Parsing mail is an ideal application scenario where a bunch of techniques combine to form one prompt. Working pipeline: first classify incoming messages (urgent, waiting for a response, maybe later, spam) - the model labels them in batches; for important ones - a brief summary of the essence and required action; for standard ones - a draft answer, which you edit and send; controversial - in a hand pile. The secret is in breaking it down into steps (marking, squeezing, drafting), and not in one “sort out my mail.” You remain an editor and do not submit anything inconsistent. Rule: the routine is broken down into a conveyor belt of simple AI steps, where a person controls the finale.