I'm not super into the whole AI boom thats occurring honestly, there are questions about the bubble and environmental impact. But, I finally got Claude Code access through work and I've been messing around with it a bit trying to decide if I like it.

So far though I'm having a bit of fun. I've guided the LLM Claude and we've created a Golang Oauth library called surprisingly bskyoauth (Not using tangled? Here is a GitHub link for all you old heads).

It comes with a demo program even that shows how to use it. Neat right?

I've been using the term "vibe-guided" since I'm not just letting the LLM run too wild (or at least trying not to, it's damn crazy), we're starting with a plan and trying to stick to it - it still goes off the rails sometimes and starts doing things I don't want it to!

Don't make changes damnit! Just create an implementation plan!!

To that end I've been messing around with GitHub's Spec-Kit. My experience so far hasn't been perfect. We still go off the rails when the LLM starts implementing stuff not explicitly in the plan. I think with some more detailed Claude directives in addition to Spec-Kit we could maybe get a bit further. I'll continue to play around with that.

My initial Spec-Kit project is a local-first Bluesky archival tool. It makes use of the Oauth tool for logging in and then will download your posts, DMs, and media, even posts you've "re-skeeted". All of it is saved locally and is search able with full text and with at: links.

I'm still on v0.1.0 and hoping to push v1 at some point, I need to make sure all our core principles are hit first. If you're brave enough feel free to check it out. Lots of work to still do but it mostly works as a backup platform already.