How to use AI as a design critic before the real review
A practical way to get useful critique from AI: what context to give it, which questions to ask, and how to sort the feedback before you act on it.

Most design problems that come up in a review could have been caught an hour earlier. The button label that only makes sense if you wrote it. The second column nobody will reach on mobile. The pricing note that contradicts the headline. You stopped seeing them because you have looked at the screen two hundred times.
AI won't replace a review with people who know your product and your users. It can give you a cheap first pass, though, so your colleagues or client spend their time on the hard questions instead of the obvious ones. The difference between useful critique and a page of polite filler comes down to how you set it up. Here is the process we use.
Know what AI can and can't judge
Before you upload anything, be clear about what you are asking for. A model looking at a screenshot is good at some kinds of critique and unreliable at others.
It tends to do well with:
- Clarity: whether the headline explains the offer, and whether labels mean what they say.
- Hierarchy: what draws attention first, and whether that matches the goal of the page.
- Consistency: mismatched wording, button styles, or spacing patterns across screens.
- Missing states: errors, empty views, and edge cases the mockup doesn't show.
It is weak at anything that needs measurement or real behaviour. Don't trust it for exact contrast ratios, font sizes, or pixel spacing; use a contrast checker and your design tool for those. It also can't tell you whether your users will understand something. It can only predict, and that prediction is a reason to test, not a replacement for testing.
Give it the context a colleague would have
If you upload a screen and ask "what do you think?", you will get generic advice about white space and call-to-action buttons. A colleague gives better feedback because they know why the screen exists. Give the model the same background in a few short lines:
- Who it is for: "Founders of small service businesses, usually on a phone."
- What the screen must do: "Get a visitor to book a 20-minute call."
- Where they came from: "A LinkedIn post about redesign costs, so they already know the price range."
- What is fixed: "The brand colours and the booking tool can't change."
- What stage it is at: "First layout. Ignore polish and focus on structure."
That last line matters most. Without it, you may get notes on icon alignment when you are still deciding whether the section should exist at all.
Ask for one lens at a time
A broad question gets a broad answer. Run several short passes instead, each with a single job. A fresh angle often turns up problems the previous pass skipped.
- The five-second pass: "Based only on what is visible first, what does this product do and who is it for?" If the answer is wrong, your first screen has a problem.
- The task pass: "Walk through booking a call step by step. Where would someone hesitate, and why?"
- The doubt pass: "What would a sceptical buyer want to know that this page doesn't answer?"
- The edge-case pass: "What happens with a long name, no data, a failed payment, or a small phone? What isn't designed yet?"
The five-second pass works well as a quick check because you can't easily do it yourself. You already know what the product does, so you can no longer read the screen as a stranger would.
Use a prompt that forces specifics
Models lean toward being agreeable, and they will often praise your work before saying anything useful. Ask directly for the format you need. This template works for most screens:
You are reviewing a design before a client presentation. Context: [audience, goal, constraints, stage]. List the five most serious problems, ordered by how much each one hurts the goal. For each, point to the exact element, explain the likely effect on the user, and suggest one specific change. Skip praise. If something is a matter of taste rather than a problem, label it as taste.
Three parts do most of the work. A fixed number forces it to rank. Pointing to the exact element stops vague notes such as "improve the hierarchy". And separating taste from problems saves you from reworking things that were fine.
Make it argue with your decision
The most useful critique is often about a choice you have already made. Tell the model what you decided and why, then ask for the strongest case against it: "We put pricing below the testimonials because buyers need trust first. Argue against this. What would we lose, and what would you do instead?"
You don't have to accept the argument. If you can answer it, you now have a clear explanation ready for the review. If you can't, you have found a weak spot before someone else did. This also works in reverse: ask it to defend the option you rejected, which is a quick way to check you didn't drop a good idea too early.
Sort the feedback before you act on it
You will get a mix of sharp observations and confident nonsense, sometimes in the same paragraph. Don't start editing straight away. Put each note into one of three groups:
- Fix now: clear errors you can confirm yourself, such as a confusing label, a missing error state, or a contradiction in the copy.
- Test or ask: claims about what users will think or do. Add these to your usability test or your questions for the client.
- Ignore: generic best practice, notes that conflict with your constraints, and anything based on something that isn't actually in the design.
Check that last point carefully. Models sometimes describe elements incorrectly or invent ones that don't exist. If a note doesn't match what is on the screen, drop it. And if a note sounds right but you can't explain why it matters, don't act on it yet. You are the one who has to defend the design in the room.
Fit it into your process, and watch what you share
Keep this small. Fifteen minutes at two points is usually enough: once when the structure is in place and once before you present. Critiquing every small change only produces noise, and it trains you to look for approval when you should be making decisions.
Before you upload client work, check your contract and the tool's data settings. Unreleased products, real customer data, and anything under NDA need extra care. When in doubt, ask the client or replace sensitive content with placeholder text first.
Then bring the results to your real review. You might open with: "We already checked the flow and the edge cases, so we'd like your view on the positioning and the pricing section." That one sentence moves the conversation toward the questions only people can answer.
Conclusion
AI critique works when you treat it as a first reader with no stake in your design and no knowledge of your users. Give it context, ask one question at a time, force it to be specific, and judge every note yourself. Try it on the next screen you are about to show someone: run the five-second pass, fix what you can confirm, and walk into the review with better questions.


