Ask for a blind review

You don't need anything special for this. Open a new chat in whatever tool you already use, so it has none of your conversation. Give it the draft and the sources the draft came from, and nothing else.

If you've been working in one long chat, ask that chat to write the reviewer's instructions, then paste them into the new one. Either way, the new chat sees what your reader will see.

That last line keeps it to a list. A rewrite buries the problems in new sentences, and then you can't tell what changed.

A reviewer can't question an assumption you hand it as background. We've made this mistake more than once. We gave a reviewer a document along with a note saying it was the right version. It was the wrong one, and nothing in that review could have caught it, because the note was background, not a claim to check.

So put every claim you want checked in a list, and say where each one came from: the customer told us, or we worked it out. Then ask the reviewer to try to disprove each one.

Keep your conclusion out of it too. If you've already written a summary or a recommendation, leave it out of the review. Once a reviewer has read your answer, its agreement mostly proves it can read.

Asking the same model again mostly gets you the same blind spots again. We once had AI count the spaces in a drone photo of a parking lot. One model said 168, then 168, then 169. That looks like confidence. A different model looked at the same photo and said 95.

Measuring the painted lines in one row against the standard width of a parking space pointed toward the higher number, but neither count was good enough to use. We set the AI counts aside and asked a person to count from the photo.

Where a task has a physical or arithmetic check, use it. Paint lines have a standard spacing. Totals have to add up. A measurement doesn't share a model's blind spots.

Disagreement is useful, too. For the Chick Springs history project we transcribed handwritten land records from the early 1800s. At least three different models read each one. A separate AI judge, blind to which model wrote what, compared their readings against the page one line at a time. It decided from the ink, never by majority vote.

The places where readers disagreed told it where to look hardest. Lines that neither the ink nor other records could settle stay marked uncertain instead of guessed, about a dozen of 313 lines across six documents.

Fix what it finds, then run a fresh review on the new draft. Each round tends to turn up smaller problems than the last, and a fix can bring in a mistake of its own. Stop when a pass comes back with only nitpicks.

Anything you add along the way goes back through. A new paragraph or a new number hasn't been checked just because the draft around it has.

A fact check asks whether the draft is right. A reader check asks whether it works for the person getting it, and the two catch different things. Run both, each in its own fresh chat.

Here's what that looked like on a weekly ad report for a client. Before it went out, a fact reviewer rebuilt each claim from the source data, and we ran it until it came back clean. A line about money we had held back was false. Spending was right on plan.

A jump in the click rate came mostly from dropping a weak ad placement midweek, so the report credited that change. Otherwise next week would have looked like a drop. A line saying a competitor had pulled back was true on one platform only.

Then a second reviewer read it as the client and named the question they would ask first, and the answer moved near the top. Four reports went through this that day, and every one changed.

Everything above works by hand. Once agents do the reviewing for you, a few more things start to matter.