Design review
Craft + UX review of a user-facing change before merge — visual hierarchy, accessibility floor, and AI-slop tells, against an approval bar. Use in the VERIFY phase for any change that touched a UI surface, parallel to /security-review. Judges by default; reports severity-tagged findings, does not rebuild unless asked.From its SKILL.md
npx -y skills add vasuag09/harness-claude --skill design-reviewAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 2 stars2 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
4.2 KB, as published. Nobody here has run it
/design-review — craft & UX gate
Goal: prove a user-facing change is crafted, not merely correct, before it merges. The
design counterpart to /harness-claude:security-review — a conditional VERIFY gate that
blocks on Blockers. It checks the build against the design brief from /harness-claude:design
and the bar in references (skills/design/references/review-bar.md + ux-floor.md).
When to use
Only when the change touched a user-facing surface (a page, screen, component, form,
email, CLI/TUI presentation). Internal/back-end changes skip — say so. Run it in parallel
with /harness-claude:review and /harness-claude:security-review when the change is both
UI- and security-sensitive.
Do this
- Scope & intent. Bound what's under review (the branch's UI diff, or a named surface).
Read the
/harness-claude:designbrief if one exists, and the project's saved design-system note (e.g..design/system.md) if present — review against the design's own intent, not a generic ideal. Apply the harness's review standard inreferences/review-bar.md. - See the whole first (squint, before line-picking): does one thing lead, or is it a
parking lot? Does it breathe, or a monotone grid? Does it look like this product? The
worst slop is compositional. Squint the rendered output, not the source — serve it over
http and screenshot (a headless browser usually blocks
file://); a picture is worth 1000 tokens. - Run the lenses independently (review-bar.md): Hierarchy · Type & color · Surfaces & depth · Composition & rhythm · States/polish/motion · Structure-reuse-content.
- Apply the floor (ux-floor.md): a11y §1 + touch §2 are hard gates — contrast 4.5:1, visible focus, keyboard nav, 44px targets, reduced-motion. Automated checks catch ~30%; the floor isn't met until you've manually exercised it — keyboard-only through the whole flow, a screen reader (VoiceOver/NVDA) over the key path, and 200%/400% zoom without breakage. Cite the specific WCAG success criterion per failure (e.g. 1.4.3 contrast, 2.1.1 keyboard, 2.4.7 focus-visible). A failure here is a Blocker, full stop.
- Score and filter. Tag each finding Blocker / Should-fix / Note. Cut false positives:
taste, a bold choice working as intended, out-of-scope/untouched lines, anything
system.mdratifies, lint/format concerns. A few high-conviction findings beat forty nitpicks.
Severity → action
| Severity | Meaning | Action |
|---|---|---|
| Blocker | Reads as generic/broken, or fails the a11y/touch floor | BLOCK merge — fix, re-review |
| Should-fix | Real craft gap a design lead would call out; still functions | Fix before merge when reasonable |
| Note | Minor | Optional |
Presumptive blockers (unless justified against intent): no focal point · size-only/defaulted typography · monotone layout · timid/competing-accent palette · harsh/fragmented surfaces · missing states · structural hacks · an inaccessible hand-rolled control where a primitive should be used · any a11y/touch floor failure.
Output
A verdict (approved / not approved) + findings. For each: what defaulted, why it
costs the user / reads generic, the specific crafted fix (the decision, not a patch),
tagged with severity. Blockers first. Judge by default — rebuild only when the user asks;
when they do, rebuild from the decision (re-derive focal point / hierarchy / surfaces), not as
patches over defaults. For a fast "make it stop looking generated" pass instead of a full
review, run the deslop mode in references/review-bar.md.
Exit criterion
No Blockers remain (or each is explicitly justified against intent), and the a11y/touch floor
passes. Then back to /harness-claude:test for any regression, or on to /harness-claude:ship.