What I Kept After Letting an AI Run My Blog

For a week I let Claude run this blog — draft posts, generate the cover images, publish them itself through a server I wired up. It makes a good demo. But when I looked at what actually earned its place, it was one small, unglamorous piece, and not any of the parts that sound impressive.

Start with what didn't matter. The auto-generated cover art looks good and solves a problem I didn't really have — a blank rectangle. The redesign was overdue, but redesigning your own site is just work, not a story. Both were me enjoying myself: useful to me, uninteresting to you.

The piece I kept is publishing. Not because an AI can write a post — that's the easy, noisy part — but because the thing stopping me from writing was never the writing. It was the publish step: finish a draft in a chat, then leave to stage, commit, push. Small, but enough that posts died in drafts. Take it away and I publish things I'd otherwise have dropped.

The one engineering bit worth copying: I put publishing behind an endpoint that commits to GitHub, and kept the GitHub credential on the server, not in the client. So any client — Claude on my laptop, ChatGPT in a browser — publishes by calling a URL and never holds a token of mine. That's the pattern; everything else is plumbing.

Here's what I keep coming back to. The appeal of "I let an AI run my X" is supposed to be the autonomy. It isn't. The autonomy is the demo. What mattered was that it removed one specific friction I'd been quietly tolerating for years. Every "AI runs my X" worth building is really "AI removed the one step that made me avoid X." Find that step. Ignore the rest.

One cost, since I'm being honest: when publishing is one sentence away, you publish things that shouldn't exist. This post replaced three I wrote that week that were really just changelogs. Easy to ship isn't the same as worth shipping — which is its own quiet argument for keeping a human in the loop, even when the loop is mostly you deleting your own posts.