The gate now has eyes. /ss-score reads the code — but some of the worst "AI-made" tells only exist in pixels, and a code-reading gate is blind to every one: a hero that doesn't actually dominate, a lower third of dead whitespace, a web font that silently failed to load and fell back to Times, two colors that look like two accents once rendered, an empty state that's a blank white void. A human catches these in half a second.

Added

• /ss-verify — the VISUAL gate. It renders the screen (dev server + headless screenshot, a static file, or an isolated component harness), the agent looks at the image, and scores the same StyleSeed gate against what it actually sees: the squint "does this look AI-made" test, whether one element truly dominates, dead/cramped space, fonts that didn't load, hues that read as two accents, contrast as rendered, optical alignment. It renders the empty/loading/error states too — the blank-void failure you can only catch by looking — then fixes and re-renders until it passes. Never claims a visual pass without a real screenshot; falls back to /ss-score if nothing can render.

Changed

• The Quality Gate is now two passes in order: /ss-score (code-clean, cheap) → /ss-verify (pixel-clean, the real bar) whenever the screen renders. /ss-build finishes a renderable screen with it. Wired into the CLAUDE.md / AGENTS.md gates, /ss-build, and llms.txt. 19 skills total.

Update: run `/ss-update`, or `npx skills add bitjaru/styleseed`.

A blanket "no scroll-linked / no parallax / no 3D / no animated gradients" was right for product surfaces (dashboards, data, forms) but got over-generalized to every surface — which made it impossible to build the exact brand pages StyleSeed points to as best-in-class (family.co, stripe.com, linear.app). Those pages are scroll-choreographed. This release scopes the rule instead of dropping it.

Changed

• Cinematic tier (DESIGN-LANGUAGE §43, PAGE-TYPES Landing): on a marketing/landing/brand page, scroll-LINKED reveals, pinned/sticky sections, the "product assembles as you scroll" move, subtle parallax, a 3D/tilt hero, animated gradient/mesh or video backgrounds, and rich hover are now explicitly allowed — with guardrails: purposeful (not jitter), 60fps (transform/opacity only), never blocks the first read or the CTA/LCP, prefers-reduced-motion leaves a complete static page, one motion language.
• The distinction the old rule blurred: scroll-JACKING (hijacking scroll speed, trapping) stays banned everywhere; scroll-LINKED (native scroll drives it, user in control) is now encouraged on brand pages.
• App/dashboard/data/form surfaces keep the restraint — the bans got a surface, they didn't go away.
• Propagated across DESIGN-LANGUAGE, PAGE-TYPES, VISUAL-CRAFT, METHODOLOGY, the CLAUDE.md/AGENTS.md gates, and /ss-score (surface-aware), /ss-motion, /ss-dial.

Update: run `/ss-update`, or `npx skills add bitjaru/styleseed`.

"Make it more minimal" is something you can just say — the model already reads plain language. A skill only earns its place where one word must move many tokens at once, coordinated, without breaking a rule — and where doing it by hand gives an inconsistent one-off. Two new skills cover exactly that; vibes stay words.

Added

• /ss-dial <axis> <direction> — turn ONE design axis up or down as a deterministic transform. Seven axes: density · hierarchy · radius · elevation · color · weight · motion. Each is a defined ramp that moves the whole coordinated token set together (density denser = spacing + padding + gap + line-height + type, all at once), respects the guardrails (8px grid, ≥44px touch, single accent, nested-radius, ≤8% shadow, font-size table), clamps at the ends (bounded, not runaway), writes the lock, and re-runs the gate to ≥ 80.
• /ss-restyle <preset> — apply a named aesthetic as a coordinate across the dial axes plus a font and one signature move: swiss · editorial · technical · warm-dtc · minimal-mono · brutalist-lite. A preset is a full identity, not a filter you stack (two presets = the mixed-personality tell). Honors a deliberate brand hex, holds the modern floor (§CC-9d), rewrites the lock, re-gates. Glassmorphism / neumorphism deliberately excluded as dated.

Why this shape

A mood word ("premium", "editorial") isn't one axis — it's a position across several → a preset (/ss-restyle). A single measurable axis with a guardrail ("denser", "sharper", "more muted") → a dial (/ss-dial). Anything the model handles from plain words stays words — no skill. 18 skills total.

Update: run `/ss-update`, or `npx skills add bitjaru/styleseed`.

The reference demo looks designed and a free-hand build often doesn't — with the same 74 rules. The gap was never the rules; it was a loop the demo runs and a free-hand build skips: lock the look → build → run the Quality Gate → fix to a floor → repeat. v2.7.0 turns that loop from prose-you-can-ignore into machinery.

Added

• /ss-build — build a screen the demo way, enforced. Refuses to write UI before the look is locked in STYLESEED.md, reads the full rules, builds with one focal point / surface-correct type / no icon-chip cliché, then runs /ss-score as a loop to ≥ 80 before presenting — with the score and what it fixed. The front door for building UI now. (16 skills total.)

Changed

• The one-paste prompt bootstraps its own install (npx skills add bitjaru/styleseed) first — because the Quality Gate is what strips the "AI look," and /ss-score / /ss-build only run when installed. Pointed at the rules-URL alone, the gate degrades to an honor-system self-check the agent skips and the lock never persists. Install-first makes the gate real.
• Landing "install" tier rewritten to state why install beats paste.
• CLAUDE.md / AGENTS.md "How to use" now lead with /ss-build (or the same loop by hand).

Update: run /ss-update, or npx skills add bitjaru/styleseed.

Full changelog: https://github.com/bitjaru/styleseed/blob/main/CHANGELOG.md

A three-domain stress test — Korean mobile health app, dark observability dashboard, warm DTC e-commerce landing, all built with zero user direction — confirmed v2.5.0 escapes the first-generation AI tells: no default indigo, no icon chips, no rainbow lists, and three genuinely different results (calm teal + Pretendard + soft / signal teal + IBM Plex + sharp + tonal dark / terracotta + Bricolage Grotesque + pill). The builders also honestly reported ~12 rule contradictions and gaps. This release patches all of them.

The 2nd-generation tells (new)

CC-9c — don't let the escape hatch become the new uniform. Ghost index numbers (01·02·03) on every section, identical uppercase-overline + big-number cards, the stock text-left/visual-right hero, the symmetric 8+4 grid — the exits from the last cliché are becoming the next one. Rule: one signature treatment per project, varied section anatomy. Enforced in the Quality Gate and /ss-score.

Contradictions fixed (found by dogfooding)

• Dark-mode elevation — light = layered shadows ≤8%; dark = tonal surface ramp + hairline borders. The gate no longer punishes correct dark-mode hairlines.
• One 8px spatial grid (stray 6px guidance removed); Golden Rule 5 scoped for mobile vs desktop.
• Radius personality → component mapping (sharp 2–4/6–8 · soft 8–10/12–16 · pill full/20–24) — one scale, not one number.
• 44px touch vs 36–40px pointer targets.

Gaps filled

• Accent ≈ semantic-color collision rule (green accent vs success green) + "completed/normal rests grey".
• Locked content/imagery palettes for product art (2–3 material tones ≠ a second accent).
• Skinless tint formula (10–14% alpha over card, light & dark).
• Desktop app-chrome type scale (dashboard h1 22–24px, hero KPI 48–64px), dense-data exceptions to the 14px floor, duration KPIs, Korean/CJK tracking guidance.
• Page-type-aware Quality Gate (static mockups/landings: States = N/A).
• Warm/consumer smart defaults (e-commerce → terracotta/coral; health → desaturated green-teal).

Full notes: CHANGELOG.md

Coherent-but-generic still reads "an AI made this." Dogfooding real product builds surfaced the next failure mode after ugliness: StyleSeed reliably stopped rainbow/emoji/pure-black, but could still leave a screen generic — the default indigo, a tight mobile type scale on a desktop, the "Lucide-icon-in-a-pale-chip" cliché repeated for every feature, an all-even grid with no focal point, and the demo layout copied verbatim. v2.5.0 forces the identity layer that made the reference demo look designed.

Highlights

• Mood step in Quick Setup — pin the aesthetic in plain words (edges → radius, feel → shadow, density → spacing/type, tone → motion), from a skin-based default you tweak ("sharper corners").
• Smart domain defaults — infers a domain-fit skin/color (Korean + fintech/regulation → Toss #3182F6). The unlocked default indigo (#5E6AD2/#4F46E5) is now forbidden as a final choice.
• Type scale by surface — a desktop/web B2B scale (body ≥16px) distinct from the tight mobile scale, plus font-pairing recommendations.
• No icon-chip cliché (VISUAL-CRAFT CC-9b) — drop the pale-chip, vary the treatment, or use numbered/typographic markers (01 · 02 · 03).
• One-focal-point rule — an all-even card grid is the machine-composed tell.
• Quick Setup is now mandatory, and the Quality Gate + /ss-score now fail default-indigo, copied-demo, no-focal, and 14px-on-desktop output (/ss-score gains a Distinctiveness category).
• Retrofit path (/ss-update Step 6) — re-do a screen built with an older, weaker StyleSeed to the new standard; report the before/after score.

Full notes: CHANGELOG.md

Theme: consistent application. Real usage showed the one-paste prompt could still produce mediocre UI (random colors, no key color, different every run). Root cause: consistency comes from constraints, and the paste path is the least-constrained. This release raises the floor of the default path — validated by dogfooding the engine end-to-end.

Added

• Quick Setup protocol — agents work in plan mode, lock a key color + motion seed with the user before building, then self-check.
• Design Lock — agents write & obey a project-root STYLESEED.md (skin, accent, radius, motion) every prompt, so decisions persist instead of drifting. /ss-setup writes it.
• Troubleshooting guide (README) + FAQ entries for "applied StyleSeed but it still looks bad / results vary."

Changed

• Stronger canonical paste-prompt — points to llms-full.txt (the full rules, not the summary) and bakes in plan-mode + key-color lock + coherence self-check.

Rules (gaps closed from real failures + dogfooding)

• No emoji as UI icons — they inject many uncontrolled hues; use one line-icon set in currentColor.
• Status color = severity, not decoration — fixed a rule that told agents to "vary status colors for visual interest" (the cause of noisy rainbow lists). Normal/OK/"보통" rows are neutral grey; color marks only the minority needing attention; same value → same color.
• Quality/score scales (good→bad, e.g. air-quality/health/life-index) now have guidance — neutral+exception or one sequential ramp, never scattered process hues.
• Favorite/saved/rating affordances convey on/off by fill (filled accent / outline grey), not gold/red.
• Enforced in /ss-review, /ss-score, the design-review skill, and .cursorrules.

Full notes: CHANGELOG.md · Update: run /ss-update or npx skills add bitjaru/styleseed

This release is about design judgment — teaching the engine not just rules, but taste: how to bias those rules for the kind of app and screen you're building, the research-backed craft that makes a component look intentional, and how to keep a whole UI coherent.

New: two axes of judgment

• APP-PLAYBOOKS.md — how to bias the rules for the app's domain (fintech, SaaS, e-commerce, social, content, productivity, health, dev-tools, …). A fintech dashboard and a social dashboard apply the same rules differently.
• PAGE-TYPES.md — how to bias for the screen type (dashboard / form / landing / detail / list / settings / onboarding). Domain × page-type = the actual judgment.

New: VISUAL-CRAFT.md — research-backed craft + Coherence Laws

The "design sense" layer, made concrete and sourced (Refactoring UI, Material 3, Apple HIG, IBM Carbon, WCAG 2.2, FT Visual Vocabulary, Butterick, NN/g, Tufte):

• §C0 Coherence Laws — one choice per axis (radius / shadow / accent / spacing / icons / type / motion), applied system-wide. The antidote to "AI-generated UI looks off": pick a corner personality and keep everything on it.
• Spacing & rhythm · typography (+ type recipe by app type) · component craft (layered shadows, the nested-radius law, buttons/cards/inputs/icons/states) · color & dark mode · data-viz · 2025–26 trends (durable vs fad).

Rules: 69 → 74

DESIGN-LANGUAGE.md gains §70 Form & Input · §71 States · §72 Accessibility · §73 Responsive & Mobile · §74 Rule Priority & Conflict Resolution.

Skills: 14 → 15

• /ss-score — score a UI 0–100 across weighted categories with a prioritized fix list. Now includes a Coherence category.
• /ss-review — new §8 Coherence checklist (one-choice-per-axis, nested radius).

Web

• /guides — 8 programmatic design guides by app type, each surfacing the domain's type recipe (typeface / hero / body / numerals) and corner personality.

MIT licensed. ⭐ https://github.com/bitjaru/styleseed

StyleSeed grew from a static design system into a full design engine with a named motion language and a live showcase. This release collects everything since v2.1.1.

✨ Named Motion System

• 5 personality seeds (Spring / Silk / Snap / Float / Pulse) × 5 contexts — spreadable framer-motion recipes
• 20+ keyword moves — tilt-3d, magnetic, glow-pulse, gradient-sweep, blob-morph, spotlight, toggle-flip, reveal-blur, confetti-pop, and more
• New /ss-motion skill — say a seed or a keyword, get the exact recipe
• One source of truth (engine/motion/library.ts) shared by the skill and the gallery

🖼️ Live galleries & showcase

• /motion — preview + copy every move
• /motion/guide — vibe-code your own motion (composition patterns + prompts)
• /interactions — toast, bottom sheet, swipe-to-delete, segmented control, like-burst…
• 11 reference builds (added food-delivery, fitness, music B2C apps) — toggle skin × motion live in the browser

🎨 Skins

• Now 7 brand skins (added Raycast, Arc) + 58 via awesome-design-md

🧭 Positioning & docs

• Reframed as a design engine — "teaches judgment, not just data"
• SEO/GEO: per-page metadata, JSON-LD, sitemap, robots (AI crawlers welcome), llms.txt
• Safer updates — UPDATE.md + /ss-update now make clear that updates are additive & reversible

🩹 Fixes

• SKILL.md frontmatter is now valid YAML → all 14 skills are discoverable via npx skills / agentskill.sh
• Reconciled doc counts (48 components · 7 skins · 14 skills · 20+ moves) and repaired broken links
• Corrected the wiki (real /ss-* skill names, working install paths) — fixes #1

Try it: https://styleseed-demo.vercel.app · ⭐ a star helps more people (and AI tools) find it!

Critical: Tailwind v4 Font Size Anti-Patterns

Documented from a real production incident. If you tokenize font sizes as CSS variables, Tailwind v4 will break your UI.

What Went Wrong

• text-[var(--text-sm)] → Tailwind interprets as color, not font-size → color: 13px (invalid)
• --text-* CSS variables → conflict with Tailwind v4's internal --text-* namespace
• Changing --font-size in theme.css → breaks all rem-based spacing (icons, nav, buttons)

What We Added

• CLAUDE.md: Font Size Anti-Pattern section with 3 ✗ and 1 ✓ pattern
• DESIGN-LANGUAGE.md: CRITICAL warning after typography rules
• Prohibited Practices: explicit rules against font-size tokenization
• /ss-lint: New check that detects text-[var(-- pattern automatically

The Rule

✓ ALWAYS: text-[36px], text-[18px], text-[13px]  (explicit px)
✗ NEVER:  text-[var(--text-sm)]                    (CSS variable)
✗ NEVER:  --text-sm: 13px in theme.css             (Tailwind conflict)