AI gave me a perfect action list. Six items were wrong.

Last week, I asked Claude to pull the action items out of a messy team meeting transcript. It came back with a beautifully organized list: owners, dates, categories. I would have sent it to the team without a second thought.

When I dug deeper, it had six problems that were invisible on first glance. A task assigned to someone who never agreed to take it. An invented deadline. A decision recorded as an action item. An issue that was raised in the meeting but never given an owner, dropped entirely. Every one of those errors was formatted as nicely as everything the list got right.

AI is excellent at producing polished work that is wrong, and the polish is what makes it so hard to catch. For a long time my answer was better prompting, then checking everything by hand. It turns out there is a better way, and it is called a loop.

Instead of one prompt and one answer, one agent does the work, a second agent grades it against a standard you define, and it goes back and forth until the output meets your bar. My list went from six errors to zero in two rounds. Then I saved the whole thing as a skill, a plain text file with my criteria and the steps, so now it runs with one command. No code involved.

The mechanics are the easy part. The hard part is the human judgment and expertise that goes in: writing down, objectively, what good looks like for your work. That is the difference between AI slop and quality you can rely on, and it is why two people can use identical tools and get completely different results.