agentsclimarketplace

Superdesign

Skill cskwork/superdesign-skill/skills/superdesign

superdesign - web/mobile UI design taken past a deterministic anti-slop gate. Use for "/superdesign", designing a new UI or landing page, redesigning an existing UI, building design tokens or a design system, critiquing a design, exploring design directions, or generating an image / icon / logo.From its SKILL.md

Install
npx -y skills add cskwork/superdesign-skill --skill superdesign

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 0 stars0 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

9.7 KB, as published. Nobody here has run it

/superdesign - intent-driven elegance

Intent -> trend-aware direction that fits -> elegant UI -> verified against a deterministic anti-slop gate. One-line CSS tweak: skip this skill, edit directly.

Core principles

  • Elegance is baseline, not padding. Polished = correct, not extra credit.
  • Read intent first. Design to the brief (kind, audience, brand, constraints), never to a default template.
  • Trend-aware != trend-chasing. Apply only trends that serve the intent. Record the trend read, dated.
  • Anti-slop enforced, not eyeballed. Every surface passes anti-slop-gate.mjs + contrast gate. Designer never self-approves.
  • Diverse by intent. Route to the aesthetic family + medium that fit; no single house style.
  • Self-contained + graceful fallback. Asset/trend tooling degrades: works without Stitch/gpt-image-2/WebSearch; missing tools -> documented placeholder + baked snapshot, never faked assets or invented data. Rendered verification is the one hard dependency: build modes render with playwright-cli (the only render driver, reference/playwright-cli.md); if it cannot be installed, STOP and ask - never substitute a headless render.
  • Verify vs ground truth. Run the static gate AND render the built surface with playwright-cli, honor reduced-motion + WCAG AA, report what passed with command output.
  • Hard stops. External publish (deploy/push/post) or destructive steps need explicit consent; ambiguous brief -> one question (non-interactive run: conservative read, assumption logged in the brief, proceed).

Mode (classify the design intent, state it in one line)

SignalModeApproach
new UI / landing / app / screen from scratchCREATEdefault loop
redesign / modernize / improve / refresh existing UIREDESIGNaudit-first (reference/redesign.md)
tokens / palette / type / theme / design system onlySYSTEMreference/design-system.md
critique / audit / review (no build)CRITIQUEreview-only, run detectors (reference/critique.md)
explore / moodboard / directions / "what style"EXPLOREdivergent, trend-heavy, no commit (reference/explore.md)
image / icon / illustration / SVG / logo (asset only)ASSETfallback chain (reference/assets.md)

Medium (web vs mobile/native) is orthogonal - decide in the loop, load reference/web.md or reference/mobile.md.

Tie-breaks (one mode wins): "critique and fix" -> REDESIGN (CRITIQUE never edits). "explore then build" -> EXPLORE; build only after the user picks. Tokens AND a page asked -> the deliverable decides (token file = SYSTEM, page = CREATE). Assets needed inside a build stay in CREATE/REDESIGN; ASSET only when the asset is the whole ask.

Default loop (CREATE / REDESIGN) - role-separated

Author-independent roles. Single surface -> inline, switch role with a fresh re-read. 2+ surfaces or a parallel asset batch -> orchestrate agents, one Designer per surface; scaffolding may fan out, look-and-feel stays deep-and-narrow.

Vault = one work dir per surface, default .superdesign/<surface>/: holds design-brief.md, trend-pulse.md, claims.md, contrast-pairs.json (start each from templates/; the pairs seed is templates/contrast-pairs.example.json). templates/preflight-gate.sh <vault> <source files> reads it - no vault, no gate. Create it at step 1.

  1. Read (brief). Infer kind, audience, vibe, references, brand, quiet constraints (a11y/regulation override aesthetics). State: Reading this as: <kind> for <audience>, <vibe> language, leaning <system or family>. Two reads diverge -> ask ONE question; else do not ask; non-interactive -> conservative read + logged assumption. Record in the vault. (reference/design-brief.md)
  2. Trend pulse. WebSearch current trend lane by default; on failure use reference/trend-snapshot.md (dated) and warn it may be stale. Reuse a same-kind pulse <=30 days old instead of re-searching (note the reuse). Keep only intent-serving trends. Record dated in the vault trend-pulse.md. (reference/trend-research.md)
  3. Direction. Set dials DESIGN_VARIANCE / MOTION_INTENSITY / VISUAL_DENSITY - scale and baseline in reference/taste-core.md. Pick official design system OR one aesthetic family (reference/aesthetics.md) OR one trend lane - never mix. Pick medium, load reference/web.md or reference/mobile.md. Done when three dial values, one system/family/lane, and the medium are all named.
  4. Build (Designer). Implement to reference/taste-core.md (always authority) + chosen family/medium. Assets via reference/assets.md. Engagement-bearing brief (Read names a primary action - sign up/buy/book/subscribe) -> also reference/engagement.md. Data-dense business app (admin/dashboard/console/internal tool) -> also reference/dashboard.md (density-first dials, app shell, no marketing hero). Enforce anti-default + reduced-motion + computed contrast. No self-approval; append claims.md per surface with a Framings: line. (agents/designer.md)
  5. Critique (independent; no design edits). Re-read reference/taste-core.md + reference/impeccable-rules.md. Enumerate every text/bg pair into vault contrast-pairs.json. Run templates/preflight-gate.sh (-> anti-slop-gate.mjs + contrast-gate.mjs) on the source, then render the surface with playwright-cli (reference/playwright-cli.md), write the ## Render block, and run templates/render-gate.sh. Log every violation. (agents/design-critic.md)
  6. Verify. Fix each violation, smallest change; re-run BOTH gates (preflight + render) until green. Report passes with output. Fresh violation loops critique -> fix; stop on green. Cap: 3 critique->fix cycles; same rule still failing -> stop, report remaining violations honestly.

Roles -> personas: build=agents/designer.md, critique=agents/design-critic.md, trends=agents/trend-scout.md, assets=agents/asset-producer.md.

Mode contract (deliverable + done-when)

No-build modes (SYSTEM/CRITIQUE/EXPLORE/ASSET): load the mode's reference file, deliver its row, skip the loop. ASSET chain: gpt-image-2 -> Stitch -> real web image/SVG -> placeholder.

ModeDeliverableVerified by
CREATE / REDESIGNsurface code + vaulttemplates/preflight-gate.sh (static) AND templates/render-gate.sh (rendered via playwright-cli) green, output reported
SYSTEMtoken file + one-screen usage exampletemplates/contrast-gate.mjs on every pair; templates/anti-slop-gate.mjs on the sample
CRITIQUEfindings report (severity, file:line, fix, verdict)detectors ran on the input; URL/HTML input also rendered + templates/render-gate.sh (screenshot-only input cannot render - note it); zero edits
EXPLORE2-4 divergent directions + one recommendationdirections genuinely differ; nothing built
ASSETasset file(s) + manifest (tier used, substitutions)links resolve; palette matches; placeholders flagged

Reference map (load only what the phase needs)

ReadWhen
reference/design-brief.mdRead: infer brief, one-line read, single question
reference/trend-research.mdTrend pulse: search, evaluate, record
reference/trend-snapshot.mdTrend pulse fallback: dated offline snapshot + refresh
reference/taste-core.mdBuild + Critique: always-on anti-slop visual authority
reference/aesthetics.mdDirection: pick <=1 aesthetic family (selection map)
reference/impeccable-rules.mdCritique: anti-pattern guardrails behind the detector
reference/web.mdBuild: web stack + layout
reference/mobile.mdBuild: mobile/native (iOS HIG, Material 3, RN/SwiftUI/Compose)
reference/engagement.mdBuild: conversion/engagement craft when the brief names a primary action (SaaS/consumer/commerce/marketing)
reference/dashboard.mdBuild: data-dense business app / admin / dashboard / internal tool (density-first overlay; no marketing hero)
reference/playwright-cli.mdCritique: the only render driver - render the built surface, then run templates/render-gate.sh
reference/assets.mdBuild / ASSET: image + SVG fallback chain
reference/redesign.mdREDESIGN: audit-first protocol
reference/design-system.mdSYSTEM: tokens, scales, theming
reference/critique.mdCRITIQUE: review-only flow
reference/explore.mdEXPLORE: divergent directions
reference/sources.mdInstall commands, sources, attribution

Final checklist

  • Mode + medium stated; brief read one line; trends pulsed + dated (or snapshot fallback disclosed)
  • One accent, one type system, one radius, one theme strategy; dials declared
  • Real/generated assets (no div-mockups); reduced-motion + WCAG AA honored
  • Engagement-bearing brief -> reference/engagement.md applied (primary action obvious, useful states, real-data proof, outcome-led CTA); editorial/portfolio left alone
  • Data-dense business app (admin/dashboard/console/internal tool) -> reference/dashboard.md applied (no marketing hero; app shell + density dials VAR 3-5/MOT 2-3/DEN 6-8; KPI north-star with deltas; tabular numerals; colorblind-safe status; per-widget loading/empty/error; tables virtualized past ~50 rows)
  • Mode contract met: build modes -> templates/preflight-gate.sh (static) AND templates/render-gate.sh (rendered via playwright-cli) green with output reported; other modes -> their verified-by row
  • Smallest change for intent; surrounding style matched; no unrequested rewrites
  • Any external publish / destructive step had explicit consent

Keep looking

Skills are one crate of 613,373. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.