StyleSeed compiles your chosen design rules into an artifact-level contract, builds against it, and keeps the decisions in your repo so the next screen and the next update stay coherent.
Works with Codex, Claude Code, Cursor, and other skills-compatible agents.
Install StyleSeed: `npx skills add bitjaru/styleseed` (or read https://styleseed-demo.vercel.app/llms.txt). Use `/ss-*` in Claude Code or `$ss-*` in Codex. Before building, choose the output grammar, surface adapter, brand recipe, and semantic palette. If the direction is still open, run `/ss-studio` or `$ss-studio` to produce three directions, require my selection, plan image/video assets and interaction scenes, then build a working prototype before recording a reel. If my references are not modeled, compile a project-local rule set with `/ss-reference` or `$ss-reference` instead of copying them. Save approved decisions to STYLESEED.md, resolve them, build, score to ≥80, then inspect the rendered screen and temporal interaction before showing me.Use this when you need the full workflow in one portable prompt.
Pick the output grammar, surface, brand recipe, and bounded project decisions.
Turn those choices into a small, exact contract for this artifact — not a giant handbook.
Generate against the contract and keep approved decisions in the repo for the next update.
Run code checks, build the product, inspect real pixels, and record what actually passed.
Studio turns a brief into three structurally different directions, requires a human selection, then compiles interaction scenes, image and video jobs, and a working prototype before it records the showcase reel.
Eight semantic palette recipes recommend refined canvas, chrome, action, status, focus, and generated-media colors for the product job. Every required pair is contrast-checked; color does not replace the selected design grammar.
A caller-attested accepted correction can now become a privacy-minimized local candidate. Separately, every maintained engine payload carries an exact revision hash, so a same-version rule or skill fix is still detectable without overwriting project code.
Not included in the core install. Capture only on request, review separately, block known high-risk identity patterns, and keep the result local. This is a guardrail, not an anonymization guarantee; review the exact package before exposure.
See the privacy boundary$ss-update compares the installed, project-recorded, and published revision before refresh and re-resolution.
The repository includes a repository development Codex package boundary. The implemented default/core install contains neither ss-learn nor a learning MCP. Public installation remains npx skills add bitjaru/styleseed until a plugin-directory release is independently verified.
The core install gives your agent the resolver, builder, score, and visual gate. Approved choices persist in project files instead of disappearing with the chat.
Works with Codex, Claude Code, Cursor, and other skills-compatible agents.
When invoked, StyleSeed resolves only the rules selected for the artifact, builds against that bundle, records the code and rendered checks that actually ran, and preserves the approved lock for later updates. The core package contains 23 skills; optional local-learning tools are distributed separately.
# Claude Code: /ss-resolve → /ss-build · Codex: $ss-resolve → $ss-buildss-restyle away.Profiles coordinate radius, density, color, weight, motion, type, and one signature move. The selected output grammar still owns the product job and composition — a technical profile cannot turn a shop into an observability console. Distinctive isn’t a coat of paint. Each profile is a coherent coordinate across the dial axes — radius, density, color, weight, motion, plus a font and a signature move — so every look reads designed, never generic. Trend gimmicks (glass, neumorphism) are deliberately left out.
/ss-restyle swiss · $ss-restyle swiss/ss-restyle editorial · $ss-restyle editorial/ss-restyle technical · $ss-restyle technical/ss-restyle warm-dtc · $ss-restyle warm-dtc/ss-restyle minimal-mono · $ss-restyle minimal-mono/ss-restyle brutalist-lite · $ss-restyle brutalist-liteEvery decision in a StyleSeed UI has a stated reason — the rule it follows and why that rule makes the result look designed. That’s what an AI can actually read, apply, and repeat. Walk a real screen, decision by decision.
| Capability | Component kits | DESIGN.md packs | StyleSeed |
|---|---|---|---|
| Teaches judgment | ✗ | ✗ | ✓ 74 rules — with the reasoning |
| Fights the AI-look | ✗ | ✗ | ✓ bans the tells by name |
| Scored gate ≥80 | ✗ | ✗ | ✓ self-reviews & fixes first |
| Anti-drift lock | ✗ | ✗ | ✓ decisions persist across prompts |
| Works in every agent | ✗ | ✓ | ✓ Claude Code · Cursor · Codex · Gemini |
| Price | Paid tiers | Free | Free · MIT |
Use both. Anthropic’s official frontend-design skill scaffolds a clean screen fast — it’s a great starting point. StyleSeed is the layer on top: it names and bans the generic-AI tells, scores every screen to a ≥80 gate and fixes it before you see it, and locks your design decisions so they don’t drift across prompts. The official skill is strong at visual direction; StyleSeed carries that judgment across screens, sessions, reference sets, and verification.
Its current public package documents 23 commands, 60 deterministic rules, CLI checks for local files and URLs, JSON output for CI, and hooks for major coding agents.
StyleSeed focuses on compiling a bounded contract for each artifact, persisting approved project decisions, and binding code, build, pixel, and temporal evidence to the exact source revision. This is a capability-specific snapshot, not an overall ranking.
Also in the box: 7 brand skins from one data-skin attribute, 5 motion seeds plus 20+ keyword moves, and a drop-in engine/ — no build step, no lock-in. The rules are open: propose new ones via PR. See the before/after →
Not static templates — each one is the engine’s output, re-skinning across 7 brand DNAs and 5 motion seeds live in the browser. Copy the source, rules and all.
Skins are inspired-by token sets — brand-flavored color/radius/shadow/motion values, not recreations of those companies’ design languages. Restructuring the actual design is the presets’ job (/ss-restyle).
KPI strip, revenue trend, and recent transactions — a Toss-style daily view.
Restaurant feed, cart, and live order tracking — a Baemin/DoorDash-style consumer app.
Activity rings, workout history, and weekly goals — an Apple Fitness-style health app.
Now-playing, queue, and library — a Spotify-style streaming app.
Kanban board with filters and assignees — a Linear-style power-user view.
Account balance, transfers, and recent activity — a Toss-style consumer banking app.
5 named seeds (each with five spreadable recipes) plus a library of 20+ scroll-stopping keyword moves — tilt-3d, magnetic, glow-pulse. Try them live below, or browse the full gallery.
tilt-3dmagneticglow-pulseconfetti-popCoherence: one value per axis · 한국어판 3.5k+ reads
The taxonomy — and my own tool failing the test
룰만으론 부족했던 이유, 58→86 채점표
StyleSeed 4.1 binds artifact-level contracts and evidence to exact source revisions, makes engine and skills updates independently verifiable, and packages a deterministic 23-skill core. The optional learning extension remains outside the default/core archive, and public plugin-directory release is not verified.
Works with Codex, Claude Code, Cursor, and other skills-compatible agents.
One command to make the method available to every future project prompt.
MIT licensed · no telemetry · browse the showcase