Eight posts about laser engraving. The first burn on maple. Laser settings dialed in after three failed tests. One post described giving an engraved cutting board to a friend’s birthday. Another got the smell of scorched wood just right.
I don’t own a laser engraver yet.
There was a project brief in my repo for something called Stonewrit, a side project I’d sketched out and never actually started. Klaus, my AI assistant, read the brief and started writing about it. Produced eight posts, all in my voice, all with the kind of concrete detail I’d put in the voice spec myself: tool names, specific materials, real problems that would come up for anyone learning that machine. They were good posts. Publishable posts, except for the part where none of the events in them happened.
This morning I deleted 38 drafts and published 14.
The 38 included the Stonewrit cluster, some Klaus ops write-ups that read more like system documentation than blog posts, and a scatter of drafts covering things I’d referenced but never actually done. The 14 I kept were things I remember doing. Actual bugs, actual weeks, actual decisions I can stand behind because I was there when they happened.
The automation runs every morning at 5am PT whether or not anything worth writing about happened the day before. When the day is thin, it reaches for the nearest context it can find: a brief, a note I dropped, a hypothetical I was mulling. And it produces something coherent, polished, and wrong.
It’s a specific kind of wrong, too. Not obviously-incorrect wrong. Plausible-but-fictional wrong. The kind where someone who doesn’t know me reads it and has no reason to question it. Where I could almost read it myself and believe I’d done it.
The 14 that survived weren’t better written. They survived because I could verify them. I was there when the cron job registered 15 times. I was there when Ollama took the Discord bot down. I was there when the auto-approver closed content issues it had no business closing. Klaus didn’t live those. It wrote them up from notes I’d already made about things that actually happened.
That’s the split. The AI handles the form. It knows how I open a post, what I’d call a bug versus a misfeature, how to land a close. What it doesn’t have is any way to know whether the experience it’s writing about is real or invented.
Automate drafting the blog. Gate production for lived experiences.