Yeah the main product I'm working on (Moxie Docs) has AI generation of mermaid diagrams, I find that it works __well enough__ but I needed some strong guardrails, quality checks, and templating to make sure it doesn't generate broken diagrams (or useless ones). First few iterations it would create a diagram for everything and it was a bit zealous on when it decided diagrams would be useful.
Would you consider introducing any tooling around that? I've been building an automation setup for keeping dev docs up to date (and eventually user-facing knowledgebase docs) but would love to chat with other developers to make sure I'm doing the right thing or meeting the right needs. For me the focus was introducing easily readable / digestible docs, including mermaid diagrams where they are useful, linking to other parts of the code / docs, etc. and then a layer on top is a basic "Q&A" feature (you could do the same in any AI tool realistically) but something that surfaces docs or generates them on demand based on where the knowledge gaps are.
Have you tried any AI review tools? I've been using / trialing a few and find they do an OK job but still seem hell bent on __finding__ and issue no matter what
Built this because I'm sick of keeping documentation up to date when everyone's using AI tools now.
It will not only generate missing docs / update / remove docs as they need alignment, but you can hook up the MCP to whatever agent is being used these days and it'll inform it of codebase conventions & relevant documentation, telling it to update docs as it find them or as it makes changes (in the same PR as related work), and reduces spend on tokens by getting things right the first time without having to grep the whole codebase for context or go through 5 rounds of "follow our conventions"