Regarding docs: Most recent projects that i looked into and tried evaluating are very focused on the shiny parts and all the nice features it got and bells and whistles.... it sometimes reads like an ad or an sales pitch.
But, I'm a sysadmin. I will be carrying operational responsibility for that thing if we decide to adopt it. I'd love to know what are the most common modes for it to break, what the cuts are where you could swap in your code, resource usage, update cycles and stability guarantees of these updates. Existence of a downgrade path (looking at you, kubernetes!). Ideally you'd know ahead if you are getting something robust where you can safely take a week off without any risks or something that needs to keep a firefighting team on-call for the rest of its lifetime.
I think a lot more needs to be done on the docs than omitting some words...
Not to jump on GPT hype too much, but a good idea perhaps for an addon targeting OSS, read over issues, PRs, scrub Stack Overflow and compile a descriptive list of the most common pitfalls for a certain lib.
It's exactly the type of writing OSS authors don't like doing, and all the information is publicly available.
The problem is a lot of it is not publicly available. E.G: I asked chat gpt to help me install Python on Ubuntu 22.04, and it failed miserably.
Why? Because 22.04 is recent and many pitfalls it comes with haven't been much documented yet.
And it also don't know what is never written, but implicitly known if you deal with a lot of beginners. E.G: people get utterly confused with *args and **kwargs in Python, because it can be used at 2 different places, and depending of those places, it does completely different things. The latter is well documented, but that the brain of people cannot grok it is not.
So chatgpt will explain the same things as most of the doc, without realizing that what it needs to do is to warn humans that they are going to be confused and how to avoid it.
Humanity has a lot of implicit knowledge.*
GPT != Chat GPT, GPT-4, etc.
It’s possibly to [“just simply”] train a GPT on whatever corpus you want.
You can train GPT 2 on whatever you want, it will give you the wrong answer for everything.
The real magic arrived with GPT3, which is proprietary, so it's fair to assume your audience understands this and imply it.