Software has huge and detailed code repositories ripe for training use. There's just enough inference in current models to remix that code in useful ways for the most popular languages.

The less popular a language, the more models struggle.

Writing, UI, and presentations have similar knowledge bases.

Outside of those, quality becomes much more hit and miss. If you ask for a recipe you may get something good, or you may get something completely inedible and random.

"Domain specific knowledge" really means "strong foundations and relevant abstractions" and LLMs just don't do that reliably.