> But why not have the developer that wrote the code guide the AI to generate the content? They know the code best.
Knowing the code and knowing how to make the code, or the interface to the code, comprehensible to another user, are different things. Just like with UIs, and the fact that an expert is not necessarily the best teacher.
Anyhow, the age of monumental feats of technical writing is past. Too expensive, and the subject is too volatile for the most part. Economics dictate that we'll have to deal with the cheapest possible docs. We already do.
Knowing the source code doesn't mean someone is a skilled communicator and expecting people who are bad at <any area> to pick out problems with LLM output in that space is a losing battle
Two major issues occur unless they have experience.
1. The developer will often have what is called "Acquired knowledge". That is information that is relevant but isn't in any of the files and the developer assumes other developers know what they know.
2. Often is the case that there is more information required that doesn't sit inside the code and is not evident to get the program to work. Quickest way to find that is to get a newbie on a clean machine to follow only the instructions.
You do realize the developers only "know the code best" because they're busy writing code all day, right?
Nobody wants to be held more accountable with less control over the result.
The moment you tell the devs to focus on working with AI is the moment their guess is as good as anyone else's what the hell is going on. You're not going to squeeze more productivity this way.
Actually at this rate, developers won't be writing code anyways but they're still in a better position to guide the AI.