Thanks! To make it work predictably, we actually tested quite a few different algorithms before landing on one that produces outputs LLMs can reliably understand.

Conceptually, we're relying on common sense assumptions about how developers structure JSX. We assume you write reasonably semantic markup where the visual hierarchy matches the code structure - no CSS tricks that make the UI render completely different from what the JSX suggests.

This let us create translation boundaries that make intuitive sense to both developers and AI models.