Check against the goal, not the to-do list

The idea

When an AI says it's finished, it usually means it did the steps. That's not the same as the goal being met. A to-do list can be fully ticked while the result still has a placeholder in it, a number from the wrong file or a link to nowhere. So check the work against the goal you started with, not against the AI's summary of what it did.

Why it works

A summary is a claim, not proof. AI tools describe their own work confidently, and what they say they did and what they actually did can differ. Looking at the result itself, one level at a time, catches what a quick read misses.

The four levels

1. Is it there? Every part the goal needs actually exists.
2. Is it real? No leftover brackets, sample names, “[insert date]” or round numbers that came from nowhere.
3. Is it connected? Links open the right page, totals match their source, names and dates match everywhere they appear.
4. Does it work? Try it the way a real person would: follow the link, fill in the form, read it on a phone.

How to do it

1. Write down the goal and the must-haves before the work starts, if you can. A finish-line plan gives you this.
2. When the work comes back, ignore the summary. Go through each must-have at the four levels.
3. Treat “should work”, “looks good” and “I think that's fixed” as a prompt to ask what exactly was checked.
4. For a second pair of eyes, use a fresh chat or a different AI with the prompt below. It has no reason to defend the work.
5. Check again after any change. A check from before the last edit doesn't count.

Here is the goal and what must be true when it's done:
[paste]

Here is the finished work:
[paste]

Don't tell me what was done. Check whether the goal is met. For each must-have, answer:
1. Is it there at all?
2. Is it real, or a placeholder, sample text or made-up number?
3. Is it connected (links, totals, names and dates match their source)?
4. Would it work for a real person using it?

Mark each PASS, FAIL or CAN'T TELL and show the evidence. Don't say "looks good" or "should work". If you couldn't check something, say so.

Worked example

Common mistakes

• Reading the AI's summary and taking it as the check.
• Checking only that things exist. Placeholders and wrong numbers exist too.
• Asking the same chat that did the work whether it's done. It will usually say yes.
• Checking once, editing, and not checking again.
• Not writing the goal down first. Then there's nothing to check against.

Goes with the $25 webinar on Tuesday, November 10, 2026: When AI is wrong, and how to tell. Its reading list is in the Library.