Genuine question: do you feel this disparity despite being diligent and checking the LLM's output, or is that not factored in?

It's also been my experience that generating an entire app/portion of code can be much easier by typing out a relatively concise prompt, but this only applies if it's the kind of code I'm

1) free from having to verify afterwards,

2) not having to fight with the LLM to get right, or

3) is coming from someone else in the form of a PR that may/may not have been checked by them first.

Having to deal with these cases has sometimes been the equivalent of using a spoon to dig a foundation, at least in my experience.

Of course checking the LLM's output and testing it and iterating on it is factored in. And yeah that verification and testing and iteration where most of the time goes, and still it ends up being significantly faster than typing everything manually, and the end result quality is often higher I believe.

To avoid fighting with the LLM to get right, it's useful to first come up with the architecture manually, run it by the LLM first (without writing any code - just talking to it, asking it to find flaws in your architecture), and once both you and LLM are happy with the plan, fire it to execute. This step removes a lot of friction (but not all of it - later iterations are still almost always necessary)