> I still think it's only good for demos and getting a prototype

Have you actually tried that? Because my bet is that if your "prototype" is a anything that is very VERY traditional, e.g. a CMS, online shop, or anything that has examples online, yes it will be quick, but if it's genuinely new, namely something NOT available out there, maybe because it is relying on the latest stack that is not yet well documented, then I bet it will also fail terribly.

Edit: I personally did, namely using LLMs to make XR demos relying on a now relatively popular framework https://aframe.io and basically it fails most of the time by proposing "traditional" HTML/CSS, missing entirely that it's 3D. Anyway, long story short, didn't work for me so curious to know if the "getting a prototype" (a genuine prototype, not a codebase starting from scratch because IMHO that's different) part is validated or just an idea.

I've only seen it do prototypes of CRUD apps, e-commerce storefronts, dashboards - vanilla stuff in other words. Get an end-to-end skeleton up and running fast; it's really good for that.

Never seen it used for anything novel so I can't refute you.