Try That UI

Use case

Review AI-generated UI before it becomes code

An AI-generated UI review asks one question: does this generated proposal match the product you already have, and is the difference worth implementing? The generator is not the review.

Why a generated screen is a weak brief

An AI design tool can hand you a polished frame with no memory of the screen it was supposed to change. Spacing, labels, and a primary action may all have moved, and the only record is the image itself. Commenting on that image tells you what the frame looks like. It does not tell you what changed relative to Current, or what you decided to keep.

Try That UI is not the generator in this workflow. You bring a proposal that already exists. The workspace holds the current product still, so the proposal has a baseline. That is the opposite of asking a model to invent a new page and then treating the invention as the spec.

A checkout button, worked through

Say a generator returns a checkout frame. The primary button now says Continue. The order summary has moved below the fold. There is no error state, only the happy path. The type is tighter than the product you ship, and a discount field appeared that the current checkout does not have.

Bring the current checkout as a URL or a screenshot. Put the generated frame beside it, and scope the review to the decision you can actually make. If the question is the button, review the button. A full-page restyle hides which change was accepted. Current stays intact so a reviewer can see that the summary moved and that the discount field is new, instead of debating the frame as if it were already the product.

If the question is what the button does on press, try that control as an isolated component prototype on Proposed. The prototype is review material. It does not run your checkout, charge a card, or become the implementation. Testing the real interaction still happens in your codebase after you accept the direction.

A finished pass on that example can be specific. Keep the clearer button label. Reject the summary move, because the total has to stay visible. Do not approve the discount field or an error state that was not in the proposal. Write that down. "Looks good" is not a decision the next person can build.

What to put on the table

The current product

Use a public URL or a screenshot of the screen the proposal is meant to replace. Current stays intact. If the proposal cannot be lined up with a real screen, you do not yet have a review. You have a concept.

The generated proposal

Bring the generated UI as its own screenshot or reference, and scope it to the element that is actually in question. A full-page restyle hides the decision. One primary action, one empty state, or one form is enough to judge.

An isolated prototype, only if the interaction is the question

If the proposal changes a control, try that control as an isolated component prototype on Proposed. The prototype is review material. It does not replay your application's behavior, and it is not the implementation. Testing the real interaction still happens in your codebase after you accept the direction.

What to inspect before you keep it

Generated UI is good at a single polished state and weak at the states around it. The review is where those gaps get named, while Current is still on screen.

What a finished review produces

Keep the proposal, reject it, or share a decision record. The record should say what the scope was, what the preview showed, and what the next step is for the person who will write the code. Approving a direction is an instruction to your team. It is not a patch, a merge, or a deploy.

Read the workflow if you want the same loop without the AI-specific framing, or the screenshot comparison if you are deciding whether a markup thread is enough. For a review that is about team alignment rather than a generated frame, use the product design review.

What this review is not

It is not a visual regression test. Chromatic and UI Verify snapshot UI that is already in a branch and diff it against a baseline on a pull request. Hexby is a visual workspace on the codebase and can hand work to a pull request. Try That UI stops at the decision, before that code exists. The comparison of those tools says which job is which, including the prices they publish.

It also will not apply the generated UI to your repository, and it will not replay your app. If a coding agent is going to implement the decision, it implements it in your project, after the review. The review does not authorize that edit by itself.

Questions

Does Try That UI generate the production UI for me?

No. This workflow is for reviewing a proposal you already have, including one an AI tool produced. Try That UI does not apply that proposal to your repository. You keep, reject, or share a decision, and implementation stays in your code.

What should I bring into the review?

Bring the current product as a URL or screenshot, and bring the generated proposal as its own screenshot or reference. The current screen stays intact so the proposal is judged against it.

Can I try the generated interaction inside the review?

You can try an isolated component prototype on Proposed. That prototype does not replay your application and does not become the production implementation.

Is an AI-generated UI review a visual regression test?

No. A visual regression test compares a new build with an accepted baseline after the code exists. This review compares a generated proposal with the current product before that code change.

Create a free account