I think this a good technique to be familiar with, although in a lot of situations I've achieved similar value by simply feeding the underlying JSON data object

I think this a good technique to be familiar with, although in a lot of situations I've achieved similar value by simply feeding the underlying JSON data objects corresponding to the intended UI state back into the coding agent. It doesn't render quite as nicely, but it is often still human-readable, and more importantly both LLM and procedurally interpretable, meaning you can fold the results back into your agentic (and/or conventional testing) development loop. It's not quite as cool, but I think a bit more practical, especially for earlier stage development.