LLMs can generate docs for code: the fact this is possible at all is often treated as the end of the story.
In reality, docs' primary failure mode is staleness, which can arise in various ways. The simplest is for the docs to record a prior version and then not be rewritten, but there can also be problems of overemphasis of the most recently developed parts, since an LLM at the end of a development session tends to over-weight whatever that session was about.
In this series I set out an idea for fixing both: pinning docs to the code they describe, and writing per-session fragments for a slower, neutral process to consolidate.