[dragonwriter]

This markup language isn't structure for prompts for LLMs, it is structure for conventional programs that need to construct prompts for LLMs.

Conventional programs using structured templates with deterministic rules to construct output is... not new.

(Jinja templates have been widely used for communicating structure to assemble conversation history, tool calls, etc., into promots for open models for a while.)

[creatonez]

It is both. The structure does actually help LLMs. Interestingly, LLMs seem to respond well to the repetition of XML-like structure (that is, the presence of closing tags), even if it's just an ad-hoc constructed language with no real schema.

[Terretta]

Not sure if changed with GPT-5, but Claude was explicitly tuned to leverage XML-like tags, while GPT-3.5 and 4 seemed to prefer Markdown-like structure.

https://docs.anthropic.com/en/docs/build-with-claude/prompt-...