Writing
Notes from running this site, then short essays on putting AI into operations without losing who decides.
Notes
This site as a product: what I built, what I didn't, and what's next
Deterministic by default, AI where it earns its place, and done building but not done running.
Dependencies we couldn't see, and the one inside the checker
If we can't read an expiry, we don't call it fine. That includes the checker's own.
The failure list that lost its history
Nothing was lost. One cap was shared, and the wrong records filled it.
When a load test used up the day's database writes
The failure looked like a choice. The logs said otherwise.
Game day 1
The site raised it. A human still made every decision after.
Essays
Has AI token usage plateaued? No. Spending has.
False on usage. Mostly true on spend, for now.
Tests passed. Nobody can explain the change.
Writing got cheap. Knowing did not.
You will not read the pile. Stop designing as if you will.
You cannot review everything. So review the right things.
AI did not invent looking it up
The model did not introduce outside help. It moved it from the search bar into the editor.
AI proposes. A human authorizes.
If the AI model is down, the folder is still the ticket.
Also on this site
- How this site is run: the heartbeat, the probe and the objectives behind the Live reliability page.