In models that support it (like Claude), how does "pre-filling" the assistant's response (e.g., starting with {) help with structured output?

We know "think step-by-step" works. But when should you not use Chain of Thought (CoT) in a production app?

How does an LLM decide which tool to call? If you give a model 50 tools, what happens to its accuracy and context window?

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