“Done” is a claim until you can check it

← Journal

Made by Lizvo · AI work method · Short read · Published October 31, 2026

When an AI tool says “done”, that is a statement about the work, not the work itself. Most of the time it is right. Sometimes it is not, and the cost shows up later, when something else has been built on top of it.

The habit below is deliberately plain. It needs no tools, only a place to write things down.

1. Decide what would count as proof, before the task starts

For each task, write one line: “I will accept this as finished when I can see ___.” The blank might be a file you can open, a page you can load, a result you can reproduce, or a number you can recompute yourself. If you can’t fill in the blank, the task is not yet specific enough to be called done by anyone.

2. Keep two words apart: claimed and checked

Keep a short list of your open items with two states. Claimed means the AI (or you) said it is finished. Checked means you looked at the evidence from step 1 and it was there. Nothing moves from the first state to the second on someone’s word.

3. Ask for the evidence in the same message

A reliable way to reduce the problem is to change the request: “Do X, and show me the file, output or link that proves it.” If the answer comes back as “done” with nothing attached, treat the item as still open and say so.

4. Check the small thing yourself

You don’t need to re-do the work. Open the file, load the page, run the one command, compare one number. A thirty-second look is often all it takes to turn a claim into a checked item, or to find out it wasn’t one.

5. Write down what you found

If a claimed item fails the check, note it where the next session will see it. Otherwise the same “done” comes back in the next conversation.

None of this makes an AI more accurate, and it won’t catch every problem. What it does is make sure that when something is wrong, you find it while it is still cheap to fix.

— Lizvo

General information only, not professional advice. See the disclaimers.