In an API inference call, what is the functional difference between the System, User, and Assistant roles?

LLMs are notoriously bad at following "Write exactly 100 words." How would you design a workflow to strictly enforce a character or word limit?

In a few-shot prompt (giving examples), does the order or diversity of the examples matter more for the model’s performance?

If you need a structured response (like a JSON object), would you rather use a "System Prompt instruction" or the model's native "Function/Tool Calling" capability? Why?