qa-verify
Skill141 starsChanged 34 days agoUnlicenseView on GitHub
Verify built UI against this project's hard rules and against the design — runs the mechanical check script, then the judgement checks a script cannot make (visual fidelity vs Figma, token naming, motion choice, semantics, contrast through every frame of an entrance, responsive behaviour including live resizes, phone overflow and dark mode), and fixes what it finds in a loop until clean. Use after building or changing any page, view, section or component, before committing, and whenever the user says "QA this", "check my work", or "is this ready to ship".
4 other projects use a version of this.
Adopted by
Most recently taken up by
- arraydude/agent-skillstheir own version, first seen Nov 16, 2026
- Markssuave/Markssuave-Agent-skillstheir own version, first seen Nov 8, 2026
- ieuanign/skillstheir own version, first seen Oct 13, 2026
- greekr4/viruagent-clitheir own version, first seen Mar 31, 2026
Other versions4 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 4
- greekr4/viruagent-cli★ 39same name
- arraydude/agent-skills★ 2same name
- ieuanign/skills★ 2same name
- Markssuave/Markssuave-Agent-skills★ 1same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository20 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- aeo-visibilitySkill
- figma-to-sectionSkill
- i18n-seoSkill
- local-seoSkill
- mobile-device-qaSkill
- optimize-3d-sceneSkill
- optimize-loadSkill
- optimize-performanceSkill
- payload-adminSkill
- payload-analyticsSkill
- payload-cmsSkill
- schema-markupSkill
- seo-auditSkill
- seo-growSkill
- ship-checkSkill
- site-migrationSkill
- supabase-authSkill
- supabase-dbSkill
The file
We split this skill into the rules it gives an agent
6 rules. 0 are common, 3 are distinctive, and 3 are very specific to this project.
0
Common
Most of the 3 edited versions we read keep these.
3
Distinctive
This author's own choices. Worth a look.
3
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 28
Every FAIL must be fixed. WARNs are judgement calls: fix or justify in your summary, do not silently ignore them.
- Distinctive: line 47
If an image looks invisible or wrong, check the container first — an invented wrapper background is the usual cause, not the asset itself.
- Very specific: line 51
Every colour/spacing/radius/type value resolves to a token.
- Very specific: line 53
No literal reached @theme inline or a Tier 2 token.
- Very specific: line 98
No horizontal overflow at any width down to 320px: document.documentElement.scrollWidth === innerWidth at 320, 360, 390, 430.
- Distinctive: line 133
Run layer 1, fix every FAIL.
“Common” is measured: kept by at least 6 in 10 of the 3 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
QA & verification
Two layers. The script decides what is decidable; you decide the rest. Never report "done" on the script alone — it cannot see a design.
Browser checks below use tools/qa/ against a running yarn build && yarn
start (yarn qa:setup once; every tool takes --url; see
tools/qa/README.md). Probes need a person's UA — the default puppeteer UA gets
the robot form, which has no motion, no cursor and no scene, and "proves" a bug
gone that a person still sees. Look at every screenshot a tool writes (Read
the image) — numbers flag, the eye decides.
Layer 1 — mechanical (always run first)
.claude/scripts/verify.sh # whole src/
.claude/scripts/verify.sh src/views/about.tsx # scoped
yarn lint
yarn build # must pass before anything shipsEvery FAIL must be fixed. WARNs are judgement calls: fix or justify in your summary, do not silently ignore them.
Layer 2 — judgement checks
Work section by section. For each one:
Design fidelity (only when a design exists)
Re-fetch the Figma node — do not QA from memory or from your own earlier summary.
get_design_context for values, get_screenshot for layout. Then compare:
- Copy — character for character. Flag anything paraphrased, shortened or invented.
- Layout — column count, flex direction, alignment, order, positioning.
- Spacing — margins, padding, gaps against the design values.
- Typography — size, weight, line-height, letter-spacing. Never assume a heading is bold; designs often use 400.
- Colour — exact values, resolved through tokens rather than matched by eye.
- Images — aspect ratio, crop, radius, overlap. No effects the design lacks.
If an image looks invisible or wrong, check the container first — an invented wrapper background is the usual cause, not the asset itself.
Tokens
- Every colour/spacing/radius/type value resolves to a token.
- New tokens follow the three-tier grammar and carry a comment naming their origin.
- No literal reached
@theme inlineor a Tier 2 token. - The cookie banner and preferences modal match the site:
--consent-*tokens point at the project's palette (not the neutral defaults once a brand exists), the type and buttons look like the site's, contrast passes light and dark, and the banner copy is still the two short lines — seeobsidian/frontend/components/common.md→ "Brand the consent UI". - A value that had to be invented because the design has no token for it is flagged to the user for design review, not quietly added.
Motion
- Everything scroll-driven, revealing, staggered or layout-affecting is a spring.
- CSS
transition-*appears only for hover/focus/discrete state, withduration-[var(--duration-*)]and a token ease. - Each primitive is the right one —
Inviewfor reveals,SpringTriggerscrub for parallax,Hoverfor hover, text through the text engine. tagis semantic on every animation component.- With
prefers-reduced-motion, content is present and readable — and the page still responds: aloop:spring orwhile (alive) await …loop not gated onuseMotionOff()restarts in the same tick forever underskipAnimationand hangs the tab (found on 6+ production sites).node tools/qa/check-motion.mjs --url …checks reduced motion and the robot form. - Motion speed doesn't follow the refresh rate: per-frame increments are scaled
by
dt(a 120 Hz phone otherwise runs them 2×) —mobile-device-qa§120 Hz. - Hover springs exist only where a mouse can hover; text reveals blur once per
line, never per letter (
optimize-performance§3).
Semantics & a11y
- One
<h1>; heading levels never skip; the outline reads as a document. - Landmarks named; icon-only controls labelled; real
button/a. - Keyboard reachable, visible focus, logical tab order.
- Meaningful
alt; decorative imagesalt="". - Contrast meets WCAG AA (4.5:1 body, 3:1 large text) at every frame of the
entrance, not just at rest:
node tools/qa/axe-sweep.mjs --url …samples contrast and target-size every ~120 ms from navigation. A first-screen fade sampled mid-way is a real Lighthouse failure — reveal with transform / clip / a mask sweep while opacity stays 1 (optimize-load§3). Its at-rest failures over a canvas are real even when Lighthouse says 100. - Target size: 24 px centres, measured on the rendered boxes.
- Overlays (menu, modal): focus moves in when opened and back to the toggle on
close, Escape closes,
aria-expanded+aria-controls, the restinert.
Responsive
This project scales the root font-size with the viewport (the adaptive grid in
globals.css), so most sizing follows automatically. What still needs checking:
- No horizontal overflow at any width down to 320px:
document.documentElement.scrollWidth === innerWidthat 320, 360, 390, 430. One over-wide section zooms the whole phone page out. - Grid/flex reflow at each breakpoint; nothing cramped or orphaned.
- Touch targets ≥ 44px on mobile.
- Navigation collapses at the intended breakpoint.
- Text stays readable; nothing clipped by an
overflow+ tight leading combo. Split headlines never break mid-word ("DA/TA") — letters grouped in awhite-space: nowrapword span. - Dark mode emulated. Any overlay or menu on a theme token (
bg-background) turns dark underprefers-color-scheme: darkwhile a fixed-colour page stays light — one menu read ~1:1 on a reviewer's dark-mode phone. Give overlays their own tokens. - Resize it live, both ways — not just fresh loads at each size.
node tools/qa/resize-check.mjs --url … [--sel "<elements that matter>"]: a real window jump, drag and widen; a DevTools device-preset switch and back (DPR, mobile and touch change together — how most reviewers check a phone); a Responsive-mode drag; tablet rotation. Each must land where a fresh load at that size does: every line "=", and LOOK at every strip — a box can match while what it draws is stale (a<model-viewer>box right, the model inside it still framed for desktop). Also by hand innext dev. Fix what it finds: sizes measured once on mount, DPR or pointer read once, a one-screen layout's scroll lock, camera framing, springs holding old targets. - Anything a phone shows changed (menus, heroes, sliders, full-screen
scenes, stills) → run
mobile-device-qa— the iOS toolbar, 120 Hz, touch and safe-area checks a desktop browser never exercises.
Architecture
- Route delegates to a view; view is a Server Component;
"use client"at leaves. - Content arrives via props/hooks — nothing hardcoded in a component.
- Async data has loading/error/empty states with skeletons.
- Components under ~150 lines, in the right folder, named export, typed props.
The loop
- Run layer 1, fix every FAIL.
- Walk layer 2 section by section, fixing as you go.
- Re-run layer 1 (fixes can introduce violations).
- Repeat until clean.
Report
State plainly: what failed and was fixed, what is a WARN you consciously kept and why, any design values that could not map to tokens, and anything you could not verify (no design available, no device to test on, iOS-only behaviour checked only in emulation). Do not report a clean pass you did not achieve.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

