AGENT SKILL

design-critique

Load before reviewing, critiquing, QA-ing, polishing or iterating on any design, even one described in words: page, screen, deck, social post, stylescape or brand asset. Renders at real sizes, scores a 12-criterion rubric, names generic-AI tells, runs name-swap and squint tests and an accessibility audit, gives at most five fixes, checks build fidelity, and revises until the target is met. Also for "is this good", "does this look AI-made" or "keep going until it beats X" (gauntlet loop).

License 许可
MIT
最近更新
2026-11-10
Stars
★ 2
Forks
0

01 前言 Front Matter

name
design-critique
license
MIT
metadata
{ "agent-attribution": "computer=vm; tool=Claude Code; version=2.1.322; timestamp=2026-11-10 03:20:00 UTC", "version-timestamp": "2026-11-10 03:20:00 UTC" }
description
Load before reviewing, critiquing, QA-ing, polishing or iterating on any design, even one described in words: page, screen, deck, social post, stylescape or brand asset. Renders at real sizes, scores a 12-criterion rubric, names generic-AI tells, runs name-swap and squint tests and an accessibility audit, gives at most five fixes, checks build fidelity, and revises until the target is met. Also for "is this good", "does this look AI-made" or "keep going until it beats X" (gauntlet loop).

02 资源文件 10

03 SKILL.md 正文

Design critique

Good critique works from evidence. It judges rendered pixels at real sizes against the brief, names each defect with a location and a fix, keeps defects apart from taste, and proves every revision with a before and after.

Defaults

Setting Default Move off it when
Evidence Screenshots of the rendered artifact, never source code or a description Only prose exists: label the critique "conceptual", say what to render, and still give provisional rubric scores with the caps that already apply (a stated AA failure caps Accessibility at 2; 4+ tells cap Distinctiveness at 2)
Web and app widths 390, 768, 1440 px, full page; add 320 for reflow The brief names devices (kiosk 1080x1920, TV 1920x1080, a specific phone)
Other formats Native size: social 1080x1350 or 1080x1920, slides 1920x1080, print as a PDF raster at 150 dpi of the final trim, video as key frames Distance media (OOH, signage): also view at about 10% scale to simulate viewing distance
Rubric target 4+ on every applicable criterion for client delivery; never average Directions for selection: 4+ on Brief fit and Distinctiveness, 3+ elsewhere
Fixes per round 5 at most, most impactful first, defects before preferences Never more
Rounds Up to 3 before showing a person The user asked for autonomous iteration; against a named real bar ("until it beats X"), run the gauntlet loop with a cost ceiling
Contrast 4.5:1 body, 3:1 large text (24px+, or 18.66px+ bold) and UI Older or low-vision audiences: aim for 7:1 body
AI tells 0-1 unrequested tells The brief asks for the style; record why
Stop Target met, or only preferences remain Two rounds without score movement: rethink the direction, do not keep polishing

Workflow

  1. Freeze the brief. Audience, job, channel, mandatories, invariants and anything already approved. If something is missing, state the assumption in one line.
  2. Render the real thing. For HTML, SVG or a URL: node scripts/render_review.mjs <file-or-url> --full-page (path relative to this skill; add --serve when the page uses fetch or ES modules, --widths 1080x1350 for fixed formats). It writes one PNG per width plus report.json; add --early 350 for a first-paint frame per width (white-flash check) and video readiness. For other formats export PNGs at native size first (Figma export, pdftoppm -r 150 -png, ffmpeg -ss 3 -frames:v 1).
  3. Look at every screenshot before reading the report. Note the first thing you see, the second, and anything broken. Then read the report: horizontal overflow, clipped text, single-word last lines, contrast failures, unmeasured text over images, rendered fonts and fallbacks, axe violations, failed requests.
  4. Run the squint test. Blur the screenshot by 6-8px (or squint). There should be one clear entry point per view, then a second and third level. An even gray field means flat hierarchy; three equal blobs mean template cards.
  5. Run the name-swap test. Cover the logo and read a competitor's name into the design. If nothing else needs to change, it is not distinctive yet: find or create one ownable device (type, color pairing, graphic system, image approach, voice).
  6. Scan for tells with ai-tells.md and log them by ID (AT-01 to AT-29). Compare against the matching pair in the AI-tells gallery.
  7. Audit accessibility for anything interactive with the 15-minute audit.
  8. Score all twelve criteria in the rubric. Every score below 4 names a defect: what, where (width and region), evidence (screenshot, measured value) and fix. Apply the rubric's caps (false claim, AA failure, overflow, placeholder copy).
  9. List up to five fixes, most impactful first. Format: [Defect|Preference] criterion: problem (evidence) -> change -> expected effect.
  10. Revise under revision discipline: scoped edits, invariants frozen, tokens and components rather than one-offs, no "while I'm here" changes.
  11. Re-render with identical settings into a new folder (--out review/round-N), compare before and after at every width, confirm invariants are untouched and no other score dropped, then re-score.
  12. Stop when the target is met or only preferences remain. Report the final scores, before and after screenshot paths, open preferences as options with their trade-off, and anything you could not test.
  13. Optional gauntlet. When the user wants it to beat a named, fetchable bar, run the gauntlet loop after the target is met: hard gates first, a blind binary critic separate from the builder, exit on a blind win or the ceiling, then re-score here. A win never overrides a failed gate.

Critique output

Write the critique in this shape, in chat or in the project's review record:

Round N, <artifact> @ <version>, widths 390/768/1440, screenshots: review/round-N/
First read (5 s): <what the eye lands on first, second, third>
Name-swap: <ownable: which device | swappable: what any rival could reuse>   Squint: <one entry point | competing: what>
Tells: <AT-IDs found, or none>
Scores: Brief 4 | Distinct 3 | Hierarchy 4 | Type 4 | Color 5 | Layout 4 | Imagery 3 | Consistency 4 | States 4 | A11y 4 | Responsive 3 | Craft 4  (min 3)
Fixes this round (max 5, most impactful first):
1. [Defect] Responsiveness: hero H1 overflows at 390 (390.png, #hero) -> clamp(36px, 9vw, 64px) -> no overflow
2. [Defect] Distinctiveness: interchangeable three-card row (1440.png) -> numbered process with real durations -> name-swap fails
Preferences offered: <option A vs B, trade-off>
Not tested: <screen reader on Windows, real devices, ...>

Defects versus preferences

  • Defect: breaks the brief, a stated requirement, a measurable threshold (contrast, target size, overflow, clipping) or a craft rule with a clear consequence (a runt in a headline, mixed icon families, an invented statistic). Fix it.
  • Preference: a taste call where both options pass every measure (warmer gray, pill or square buttons, a two- or three-line headline). Offer it, label it "preference", let the client choose.
  • A preference becomes a defect when it is the client's stated requirement or a locked brand rule. Never present taste as a rule.

Decisions that change the answer

  • Stage. Exploration critiques direction (Brief fit, Distinctiveness) and ignores polish; delivery critiques everything at the 4+ target.
  • Requested style. Glass, gradients, neon or symmetry asked for by the brief are not tells; judge the execution. python3 ../../library/scripts/atlas.py show style.<slug> prints the style entry's path; read its "Clichés and AI tells" and "Accessibility notes" sections.
  • Channel. Web and apps: three widths plus keyboard and screen reader. Social and OOH: legibility at thumbnail or distance and safe areas. Print: final trim, bleed, minimum type size (7-8pt for captions as a common floor). Motion: key frames, reading holds and a reduced-motion version.
  • Audience. Older adults, children and low-literacy audiences raise the text size, contrast and target-size bar; expert users can take higher density. Query the audience entry: python3 ../../library/scripts/atlas.py query "older adults".
  • Script and market. The single-word-last-line check is whitespace-based; for Chinese, Japanese, Thai and other scripts without spaces, check line breaks by eye. Check RTL layouts mirror correctly.
  • Locked assets. When the logo, palette or copy are fixed, critique the system around them and say which defects the locks cause.

Common failures

  • Critiquing from code or a description -> render first and look at the pixels.
  • Checking only 1440 px -> most breaks happen at 390 and 768.
  • Twenty fixes in one pass -> five, ranked; the rest go to the next round.
  • Averaging scores -> the minimum governs delivery.
  • "Clean" mistaken for finished -> run the name-swap test; clean and interchangeable scores 3.
  • Fixing a runt with a non-breaking space at one width and causing overflow at another -> re-render every width after every change.
  • Treating the report as the verdict -> it lists leads; the screenshots decide.
  • Endless polishing -> obey the stop rule and show a person.
  • Claiming approval -> say "ready for review"; only a named person approves.

References

  • Rubric: twelve criteria with 1-5 descriptors, caps and a score sheet. Read before scoring.
  • Cinematic addendum: blocking checks, signature, media integration and motion criteria, and the 8+ average, none-below-7 stop rule for cinematic landing pages and heroes.
  • AI tells: 29 named tells with what they look like, why they happen and the fix.
  • Accessibility audit: the 15-minute pass (axe, keyboard, focus, zoom and reflow, contrast, targets, motion, VoiceOver and NVDA).
  • Gauntlet loop: after the rubric target is met, builder and blind binary critic per piece against a named, fetchable bar captured at the same widths; gates first, copy guard, cost ceiling, round log, single-context fallback. Adapted from RoboNuggets' gauntlet-loop (CC BY 4.0).
  • Review panel: art-director, accessibility-auditor, copy-editor, brand-guardian and frontend-engineer subagents, score merge into the rubric, conflict ladder, chat fallback.
  • Build fidelity and compare_build.mjs: pixel and computed-style diff of approved design vs build, thresholds, masks, pairing with Core Web Vitals.
  • Revision discipline: scoped changes, invariants, before and after evidence, stop rules.
  • render_review.mjs: Node 18+, no bundled dependencies. Install Playwright and axe-core in the working project (npm i -D playwright axe-core && npx playwright install chromium) or point PLAYWRIGHT_MODULE, PLAYWRIGHT_CHROMIUM and AXE_SCRIPT at existing copies. --exclude SELECTOR skips deliberate controls or third-party embeds; --early MS saves a first-paint frame and reports the painted background; --video-wait MS waits for videos before measuring; the report lists video readiness and errors, broken images, failed requests and console errors; --strict exits 2 on blocking findings for CI.
  • AI-tells gallery and its README: 18 tells rendered generic versus better across 18 industries and styles.
  • Quality benchmark: FORM and TIDAL calibration briefs with generic and wrong controls. Compare before you score.
  • Stylescapes: twelve excellent, distinct directions to calibrate what a 5 looks like outside one aesthetic.
  • Deep process for client projects: design taste, website visual review, website accessibility, browser QA, media quality evaluation, delivery and stage handoff.

Catalog IDs covered: design-taste, web-visual-review, web-accessibility, web-browser-qa, media-quality-evaluation, delivery

04 同仓库更多 Skills

newmindsgroup
newmindsgroup/visual-design-kit

web-design

Load before planning, designing, auditing or launching any website, landing page, cinematic hero, microsite or above-the-fold layout, or a prompt for Lovable, Bolt or v0 to build one, including regulated (prescription drug, health, finance), multilingual or market-specific launches. Covers page types (home, product, pricing, about, blog), conversion and ethical persuasion, responsive and fluid layouts, Core Web Vitals, SEO and structured data, localization and language-law requirements, stack and CMS choice (Astro, Next.js, WordPress, Webflow, Framer) and launch QA.

★ 2
newmindsgroup
newmindsgroup/visual-design-kit

ux-design

Load before designing or reviewing any flow, form, onboarding, navigation or content for real users, especially low-literacy, older, neurodivergent, child or other specific audiences. Covers research and usability tests, heuristic reviews, personas and jobs to be done, journeys and service blueprints, information architecture, UX laws, SUS, SEQ and HEART metrics, plain language and reading level, and microcopy and error messages.

★ 2
newmindsgroup
newmindsgroup/visual-design-kit

ui-design

Load before designing or specifying any interface screen or component, even a quick one: forms, tables, navigation, dashboards, onboarding, auth and passkeys, pricing, checkout, search and filters, settings, empty, loading and error states, notifications, consent and age-appropriate flows, and AI chat and agent interfaces, all to WCAG 2.2 AA. Also picks React UI libraries (toasts, drawers, command menus). Gives component specs, state matrices, sizes, density and keyboard and accessibility behavior.

★ 2
newmindsgroup
newmindsgroup/visual-design-kit

typography

Choose, pair, size and set type for brand, web, app, print, slides, signage and motion: modular scales, fluid clamp() type, measure, leading, tracking, OpenType features, font licensing and web font loading, and multiscript setting for CJK, Arabic, Hebrew, Devanagari and Thai. Use when selecting or pairing fonts, building a type scale or type tokens, fixing weak hierarchy or bad line breaks, setting long-form text or numbers, loading web fonts, or localizing type for another market.

★ 2
newmindsgroup
newmindsgroup/visual-design-kit

reference-reconstruction

Reconstructs the observed interface and behavior of a supplied app, site or screenshot set using a screen index, measurements, state and command maps, reusable components and matched-condition comparisons. Use when asked to reverse engineer, clone, reproduce or match a reference product, then adapt it if requested.

★ 2
newmindsgroup
newmindsgroup/visual-design-kit

print-packaging-social

Load before any print, packaging, label, barcode, social, out-of-home, slide-deck or email design question, even a quick one, and before making finished collateral files. Specs: bleed, safe zones, 300ppi, CMYK, PDF/X preflight, dielines, barcodes, social formats with 9:16 safe areas, billboard letter heights, 16:9 decks, bulletproof email. Renders token-driven brand books, sell sheets, cards, carousels and decks to PDF and PNG with a Canva editable twin.

★ 2