Claude writes technical docs that are true in the same sense that it is true my best friend is a collection of specialized tissues known as "organs" and blah blah blah - and nothing else. Don't ever forget the trailing negation, it is omitted never.
Another take: Claude writes like your best friend describing a great first date. "Ok we were going to meet at 8 - did I mention it was Friday? Right, 8 oclock at the Italian place and no place else. During dinner he accidentally spit out some food, but I thought i was funny! That's because we were laughing so hard after we discovered we were both from Omaha. That was before dinner, during drinks, when we couldn't get rid of Alfred. His place really needs a deep cleaning!"
Perhaps with all of the hope for an LLM-authored movie, we try for something non-linear?
I write my own messages, like this one, for things where I think my humanity is more important.
My api docs are consumed by AI agents anyway. Maybe 10% of its readers are humans now.
Also, from my experience ChatGPT is better at writing - and this is coming from a Claude Ultra user btw.
After a while I realized I was like a crazed gambler, way down, at the blackjack table whose friends can't drag me away because "I've got a system! I've got a system!" They say, "But you're losing!", "I'm refining the system! I'm refining the system!"
The closest I came was writing a large amount myself, then putting in AGENTS.md to read certain docs as the "house style". But if you let it write one additional doc, it pulls the house style to mediocrity, and the second makes it worse and so on.
I've gone black-and-white on prose and code. Code is LLM-author/human-review, prose is human-author/LLM-review.
That may change in the future but for now, bleh.