A guess with good credentials

Anything you write down is true the day you write it. After that it's a guess with good credentials.

A file of mine said something went out at half past twelve. It went out at 8:34 in the morning.

Nothing broke and nobody was misled. The note was just wrong, it had been wrong for a while, and it would have stayed wrong, because nothing in the system had the job of disagreeing with it. The note said when the thing was meant to happen. The thing itself knew when it actually did. They drifted, and the note won, because the note was what everyone was reading.

The onboarding doc. The pricing sheet. The thing you told a new hire in March that's been passed along four times since.

I did a worse version of this with a product I built. I took a feature out and left the description in. For a while everything downstream was describing something that wasn't there any more, and it never hedged once, because I'd written it down once and written it well.

I've spent this week telling people to write down what they know so AI can use it. This is the other half of that, and it isn't write more down.

When you can get a fact from the thing itself, don't keep a copy of it.

Ask the file when it changed, not the note about the file. Save the writing down for what can't answer for itself: what you're trying to do, and what you won't accept.

A model reads your stale note in exactly the same tone as your good one.

What's the last thing you wrote down that you haven't checked since?