@david_chisnall this very much matches my experience. It's harder to second guess everything when it doesn't work at runtime than it is to just write from scratch and think about the bits I don't know as I go.

The made up wishful thinking "yes it would be nice if the API worked like that wouldn't it?" stuff more than drowned out all the utility of generating code where I'm not particularly familiar with the task.

I probably did write more test cases using LLMs than I otherwise would though.