We've been putting off the chore of writing technical docs since our team could handle all customer questions directly. However, since releasing an AI tool in the system, users have started asking it system questions as opposed to just data questions like it was designed for. It couldn't answer how-to questions, so, users got confused.
We didn't have a direct way to teach the Ai about our system since our internal docs were thin and the codebase large.
We ended up defining some good documentation practices in best-doc-practices.md, passed it to GPT-5.5, and asked it to write system docs for every module in the system (hundreds), handling 5-10 at a time for context limits.
To our surprise, it worked very well. The tool had read-access to all relevant server/client/database files, so the output was accurate and aligned to what we defined in the best-doc-practices.md.