I also have not had much luck with LLMs when it comes to anything with substantial complexity.

Where I’ve found them best is for generating highly focused examples of specific APIs or concepts. They’re much better at that, though hallucinations still show up from time to time.