Now

Big questions used to go to four AI models at once, and I'd let them argue it out.

Now it's two. The other two kept needing me to sign them in, and a council you have to log in one member at a time isn't a council. It's a register.

The two that stayed have rules, like any committee worth sitting on.

Rule one: nobody gets an opinion until they've read the actual thing. Not my summary of the thing. The thing.

Rule two: I don't say what I think first. One plan arrived with its author's own list of worries attached, and the council spent the whole meeting politely discussing that list. Keep the worries back and they found twenty-seven problems. The author had spotted three.

Rule three: someone has to disagree. A third model sits in with one job, which is proving the other two wrong. It is the most useful member and the least invited to drinks.

Things the council has taught me this month. Point at a problem and the fix will land exactly where you pointed, leaving the identical problem in the next room. "Done" can mean "I stopped checking". And if one isn't sure it saved your email, it'll helpfully write it again. And again.

Meetings are shorter. Nobody brings slides.

Currently impressed by: a machine whose whole job is to disagree with me.

Currently failing at: not overruling it when it does.