I'm fairly pessimistic about LLMs, but I do find they're a good research tool. If I can accelerate finding answers amid the slurry of Confluence, Slack, assorted other intranet places and the web, I can get back to thinking about the design and programming.

I have a user CLAUDE.md that tells it not to presume to modify any files, and a bunch of internal MCPs for our logging and monitoring tools, documents, git repositories and such, and I've been dabbling with writing skill files so it can access some other internal things.

So I can basically say "this thing isn't working because x, and it's probably related to y not having z" and it'll go off and find info buried in the corporate abyss.

Then other people yeet wildly awful LLM generated PRs over the fence and I lose my sanity from basically having fight a constant battle against a slow rot of software quality that will bite us later.