In practice, the LLM, treedocs and I all work together to maintain it. treedocs finds missing and extra paths, and missing descriptions. The LLM writes descriptions. git makes sure I never forget to run treedocs sync. And I always have the ability to easily correct incorrect descriptions by the LLM, or add my own special instructions.

Sure the LLM can do all of that but it's a waste of a lot of tokens, and in my experience it often makes incorrect assumptions or doesn't realize certain details.