A website design review skill for your AI agent

A website design review skill gives an AI agent a reusable method for critiquing a page. Instead of repeating a long prompt after every build, you keep the review criteria and reporting format in one place. Each review still needs its own audience, task, and current screenshots.

The PageJury skill below focuses on visual hierarchy, messaging, credibility, and mobile layout. It asks the agent to connect findings to visible evidence and separate observations from assumptions. It is a review aid, not a source of human research or a guarantee that a website is ready to launch.

What belongs in a design review skill?

Put the stable method in the skill and the changing context in the request. The method should say what the reviewer needs to inspect, how to distinguish a finding from a preference, and how to deliver a useful result. The request should name the actual page, audience, and visitor goal.

This download follows the basic Agent Skills format: a folder containing a SKILL.md file with a name, description, and instructions. The skill’s name is website-design-review. Discovery and installation locations depend on the agent you use.

Its scope is deliberately specific. A source-code security review, a search-ranking audit, and a visual design critique need different evidence. Calling all three a “website review” makes it too easy for an agent to produce confident findings about things it never tested.

How to use the downloadable skill

1. Download the file and read it. Create a folder named website-design-review containing the file named SKILL.md.
2. Place that folder in a skill location supported by your agent, following that product’s current instructions. The download does not install itself.
3. Ask the agent to use the website-design-review skill and supply your brief and screenshots. If skill loading is unavailable, attach the file as review instructions.
4. Check that the response names the evidence inspected. If the agent cannot see the screenshots, switch to a capable environment or limit the task to copy.

Use the website-design-review skill on these screenshots.
Audience: independent tutors.
Task: understand the scheduling product and try a sample booking.
Evidence: mobile and desktop screenshots of version B.
Keep: the calm illustration style and current brand colors.
Review only. Give me the most consequential findings first.

Before saving new screenshots, check whether they contain private information. Use a public page or appropriate sample data when possible. The skill itself does not add a browser, vision capability, or permission to send a page to another service.

Expect a decision record, not a score

The skill asks for up to five findings. Each includes a location, evidence, likely effect on the task, proposed change, and a way to verify it. It also asks what to keep and what was not tested.

For example, “The pricing footnote wraps beneath the sticky button in the supplied mobile screenshot” is an observable issue. “This makes visitors distrust you” is an interpretation that needs a separate check. Both may be worth discussing, but they should not carry the same certainty.

The skill does not automatically rewrite the site. If you later ask your agent to make changes, choose the findings to address and preserve the original brief. Compare the next screenshot against the same task using the revision prompt.

Adapt the rubric to your product

For a booking page, focus on availability, the meaning of the primary action, and what commitment follows it. For a portfolio, focus on the work a potential client needs to understand. For a developer tool, focus on the explanation and evidence a technical evaluator would expect.

Add those criteria to your review request before adding permanent rules to the skill. A problem found on one page does not mean every future page needs the same solution. Keep a small collection of example reviews so you can notice when the skill’s advice becomes generic or starts assuming facts it cannot see.

Try the method on a page from PageJury’s fictional practice gallery. These examples are for practice; they are not customer results or evidence of improved conversion.

Add human feedback at the uncertainty

When the unresolved issue is what a visitor understands or feels, turn it into a neutral question. “What do you think happens after selecting this button?” is more useful than asking a model to impersonate ten customers and average their opinions.

PageJury is designed to add voluntary human first impressions of a saved mobile screenshot. Explore the example report and pilot details. The skill works on its own; using it does not submit anything to PageJury.