Every week I’m going to write down what we built. “We” is me and my AI agents. Mostly that’s Friday, my chief-of-staff agent, plus a handful of others that each run one job.

This isn’t advice. It’s a build log: what shipped, what’s still rough, and what broke. This first one covers August 31 through September 6. It’s a short week, and most of it is about something that broke.

My AI made up a to-do item

Friday reads my meeting transcripts and pulls out the to-do list for the week. Who said they’d do what, and for whom.

This week it pulled from four meetings. One item on the list was something nobody said.

It wasn’t random, either. It was plausible. Friday had borrowed it from things it already “knew” about me from other work and set it down next to the real items like it belonged there, with the same format and the same confidence as everything around it.

I’d already treated it as real before we caught it. That’s the part that bugs me.

The fix: make it show its work

So we added a second pass. After Friday pulls the list, a separate check goes back to the transcript and re-reads it, item by item. Every to-do gets marked one of two ways: CONFIRMED, with the exact quote from the meeting, or NOT FOUND.

We also put a hard rule on the first pass. It pulls from the transcript and nothing else. No filling gaps from what it remembers about me.

Then we ran it on the same four meetings.

23 to-dos came back confirmed, each one with its quote: 12 from the first meeting, 9 from the second, then 2 and 1 from the last two. The made-up item showed up zero times in the transcript. It got flagged NOT FOUND, which is exactly what it was.

What the check proved

When an AI makes something up, your first instinct is to stop trusting the whole thing. Every item on the list goes under suspicion.

The verification pass is what let me sort it out. The bad item came from one of the four extraction passes. The other three were clean. Without the check, I’d have been guessing which parts to trust. With it, I had a quote for every real item and a flag on the one that wasn’t.

If you’re handing any kind of note-taking or task tracking to AI, that’s the setup I’d copy. Don’t ask it to be careful. Make it prove every line against the source, and make “I couldn’t find this” an allowed answer.

Friday doesn’t hand me anything as fact anymore unless it can point to where it got it. No quote, no spot on my list.

What’s still rough

A transcript check catches an item nobody said. It can’t catch an item somebody did say that Friday misread. That one’s still a human read.