They train it with XML even the system prompts that Claude reads are formatted by it.

I haven't done it professionally, but my understanding is that this kind of work is much more in the second category, where you have to understand the closest approximation to what you want that the LLM can reliably produce or the training won't work at all.