These little demos of toy UIs are cute but humans can still hold the context of a whole codebase in their heads.
We're not quite at the level (yet) of feeding a whole codebase to an LLM and making it add features or make changes while understanding the big picture of the problem being solved, being consistent with the design principles and coding style of the overall existing codebase. And I'm not even talking about creating complex UIs where performance matters
Another issue I've ran into a lot is staleness of the knowledge the LLM was trained on, a lot of libraries and frameworks get really frequent, quite often breaking changes, and LLMs have a cutoff date.
Try using ChatGPT for something like Godot's GDScript, it will always try to use old Godot v3 style scripting because that's what it's been taught, and the whole documentation for Godot v4 is not something small enough to just fit into context
Maybe this would be a better fit for some agent type workflow where it can decide what to lookup from the documentation and then retrieve it, but it also needs to know and decide what to look up and how. There is still a lot to figure out