Get an honest AI review

Ask the session that built your page whether it's any good, and you'll usually hear that it is. It wrote the page. It remembers every reason it had for every choice.

So the review comes from somewhere else: a fresh AI session with no history of the build. The session that built the page remembers every reason it had, and it leans toward its own work. A fresh one sees only the page and the standard you give it. This lesson is about setting that up so the review is honest, and about what you do with what comes back.

The order matters more than the prompt. Write your own critique first, against the brief. Save it with the date, so it's on record that you decided before the AI did. Then open a fresh AI session with none of the build conversation in it. Give it the standard and the build. Then put its findings next to yours.

What you give it is the standard: the brief, the positioning, the art direction, the claims list, and the build itself. What you withhold matters as much. Keep your own critique back, because the point is two independent reads you can compare. If it sees your list, it has something to agree with. Keep the build conversation out, which is the reason for a fresh session in the first place. And leave out anything that tells it what you're hoping for, like I think this is nearly done, or I made this. Every hint makes the easy answer easier.

Here's the shape of the prompt I'd use. It names the files that count as the standard. It asks for each failure with its location, the line it breaks and a fix. And it limits findings to the brief and the truth. That limit is there because a reviewer asked to find problems will usually find some, even when the work is sound. If you don't narrow it, you get a list padded with preferences. On anything a client will see, I'd also run the same prompt past a model from another company, and check anything only one of them flags.

Now reconcile. The AI's list doesn't get obeyed. It gets sorted, and you're the one sorting. Some findings are correct, and they go on your fix list. Some point at a real problem but blame the wrong cause, so you fix the cause. Some aren't problems at all. Some are taste presented as a failure. And then there's the column that teaches you most: what it caught that you missed, and what you caught that it missed.

Here's what that looks like for the fictional bakery. The faint captions finding is right: the builder added a gray that wasn't on the design's list, so the fix is a color from the list. The testimonials suggestion is the AI filling a gap with something nobody has cleared, so it's a no until the owner says which reviews can be quoted. The empty space is taste, because no line in the brief asks for less of it. The menu finding has the right feeling and the wrong cause: the builder tidied the approved label, Today's bread, into Our range. And the hidden Order ahead button on phones is mine. The reviewer missed it. Write every miss in your AI log. Over a few projects, that log tells you where to trust it and where to look harder.

The reviewer finds things. You decide what they mean. Write yours first, review in a fresh context, and never let the builder grade its own work.