Mainly to narrow the problem space: there's a lot of logic in our backend that we call "post processing" which is cleaning up the code after the AI generated it to fix hallucinations. For example, the AI often gets import statements wrong or misspells icon names.

It's pretty interesting because every model seems to introduce a new set of hallucinations, so this is a problem that requires a lot of maintenance. We have this internal mantra to "use AI as little as possible" especially it's a problem that can be solved deterministically.

Also, because we are more focused on PMs, designers, and product leaders the code is largely an implementation detail. What they care about is being able to visually communicate their ideas, and React just happens to be a great way to do that because LLMs are great are outputting it + we have a pipeline to render it. (We are React developers ourselves.)