Show HN: XML, Markdown, or JSON: Which gives LLMs the most reliable boundaries?

I've used all three and I think markdown is the most effective. I build an AI toolkit that generates TOON for datasets and that works remarkably well too.