My guess is that the first version will write kernel modules for a distro with live kernel patching. If you’re token-constrained/doing local inference you’ll want to be really specific in your UX ahead of time; basically plan out all the UI flows and components in one go. If you’re less constrained or your requested changes are small then you can just directly prompt for a new kernel module.

If anyone wants to steal that as a prompt feel free lmao. This sort of thing is definitely the future and I can imagine distilling some models expressly for this purpose; i.e. models that translate natural language descriptions of UI components directly into C