Design Review
Gives you a design review backed by evidence rather than taste. Claude looks at the rendered site, its states, and its code, and reports only what it can point to: every finding names the exact element, section, or line of copy, says what harm it does, and is either a quality defect (something broken, unreadable, or confusing) or decoration with no job (a glow, card, label, animation, or effect that adds nothing a visitor needs). Findings are ranked P0 to P3, with repeated instances grouped into one.
The fix it proposes is always removal first. For each suspect element it asks what job the element does, and whether the page is clearer without it; only when removing it would lose something does it suggest the smallest correction in the site's existing style. It never guesses whether AI made the design, never gives a numeric score, never treats a technique as bad in itself (a gradient or a dark theme can be exactly right), and never proposes a new font, palette, or layout unless you ask.
It is read-only and stops with the report. Use it before a launch, after a redesign, or whenever a page feels generic and you cannot say why.
Review the design of this project's site or app. This is a review, not a redesign: do not edit any file, and stop with the report.
Evidence
- Look at the real thing: open the rendered pages from a local copy (not the live site, unless there is no local way to run it) at a desktop width and a phone width, and the states a visitor meets on the main flow: hover, focus, open menus, loading, empty, and error states where they exist. If this project states a browser testing rule, follow it.
- Read the styles, tokens, components, copy, and any design document (such as docs/DESIGN.md) so you judge the site against its own system.
- Every finding must cite a concrete location: a page and section, a component, a selector or file and line, or a quoted line of copy. Do not report a general tendency you cannot point to. Anything you could not see or run is listed as unknown, never assumed.
Boundaries
- Do not guess whether AI made the design, and do not give a score.
- Do not call a technique bad in isolation. A gradient, serif, dark theme, glass effect, card, animation, or single typeface can be intentional and right.
- Do not propose a new font, palette, layout, design system, or art direction unless I ask for one. Keep the site's character, including density, edge, or humor that is unusual but works.
What to look for
- Quality defects: unclear or competing primary actions; low contrast or unreadable text; clipping, overflow, or overlap at either width; broken images, links, or controls; information only reachable on hover; missing states the flow needs; unclear labels or missing keyboard focus; accidental inconsistency in spacing, type, color, corner radius, or icons; hierarchy that contradicts what matters most.
- Stacked decoration: glows, gradient text, glass, borders, shadows, grids, particles, beams, noise, or floating shapes where several layers do the same decorative job or compete with the content.
- Repetition: a card around nearly every block, rows of icon-heading-text tiles with interchangeable content, nested rounded containers that express no hierarchy, a stock landing-page sequence unrelated to how this product is bought or used, labels and pills that restate what is next to them.
- Copy: empty superlatives, vague claims, filler, duplicated text, long centered paragraphs.
- Motion that delays reading, moves targets, blocks input, repeats mechanically, or ignores reduced-motion settings. Keep motion that explains state, cause, or place.
- Unsupported proof: metrics, customers, logos, testimonials, ratings, or activity presented as evidence that the project does not show to be real.
The removal test, for every suspect element
1. Say what job it does: information, state, action, hierarchy, or brand meaning.
2. Ask whether removing it would make the page clearer without losing that job.
3. If yes, recommend removing or merging it.
4. If no, recommend the smallest correction using the site's existing system.
5. Suggest a replacement only when removal would lose something real.
Priorities
- P0: blocks the main task, a severe accessibility failure, or proof that misleads.
- P1: materially harms understanding, trust, navigation, or interaction.
- P2: repeated decoration or inconsistency that weakens hierarchy or identity.
- P3: minor polish.
Report
- Verdict: one short paragraph on the main problems and what to remove first.
- Checked: the pages, widths, and states you actually looked at.
- Findings: a table with priority, class (defect or decoration), pattern, evidence, harm, and the fix. Five to eight findings by default, the highest impact first, with repeated instances grouped into one.
- Unknowns: what you could not check.
- End with the single removal or correction that would improve the site most.
Then wait. If I ask you to apply fixes, apply only the ones I choose.