★ Open source · github.com/yizhiyanhua-ai/fireworks-design→
Open-source · Claude Code Dynamic Workflow

One prompt in.
A studio's worth of directions out.

fireworks-design replicates ClaudeDesign — it fans out 6–8 distinct aesthetics in parallel, judges each across 6 design dimensions, synthesizes the best, then adversarially refines to a single shippable page.

Star on GitHub See 24 live pages →
~ in claude code, just say
Install fireworks-design into this project — repo is yizhiyanhua-ai/fireworks-design

then type /fireworks-design + your request →

24LIVE PAGES
18AESTHETICS
6JUDGE DIMENSIONS
∞MODEL-AGNOSTIC
Why teams use it

A design studio in a single workflow.

Not a prompt that hopes. A pipeline that ships — explore, judge, refine, polish.

🎛️

Explore 8 directions at once

Every run fans out 6–8 fully-committed aesthetics in parallel. You see a studio's worth of options, not one lucky roll.

⚖️

A panel judges every dimension

Six independent critics score hierarchy, type, color, motion, craft, delight. The winner is earned — with a paper trail.

✅

Ships accessible & polished

WCAG AA, responsive 375→1280+, reduced-motion, no dead links — verified in headless Chrome before it reaches you.

🔓

Yours, open, model-agnostic

One MIT file. Bring any model. Runs in your Claude Code — no subscription, no upload, no lock-in.

The philosophy

Don't ship the first draft.
Ship the verdict.

Every other tool hands you one generation and calls it done. fireworks-design runs a whole design studio in parallel — and only the page that survives exploration, judging, and adversarial refinement gets to ship.

Designed to be wrong in parallel —
so the final page is right.

A single model output is one draw from a distribution: its taste, mood, and reading of your brief are locked into that version, and its ceiling is capped by variance. We don't try to draw better. We draw many times, on purpose, and let independent critics throw out the losers.

“ Even experienced designers have to ration exploration — there's rarely time to prototype a dozen directions, so you limit yourself to a few. ”
— Anthropic, Introducing Claude Design · the shared root of the idea

Same idea. Open-sourced. Model-agnostic. Yours.

01 / BREADTH

Explore like you're infinite.

6–8 fully-committed directions, every run. One prompt gives you one mood, one taste, one lucky roll — we give you a studio's worth, then keep only what earns its place.

02 / VERDICT

Judge, don't hope.

Every direction scored 1–10 across 6 independent dimensions — hierarchy, type, color, motion, craft, delight. The winner isn't a feeling; it's a number with a paper trail.

03 / MODEL-AGNOSTIC

The model is the intern. The panel is the art director.

Inherits whatever model you run it on. A weaker model explores more crudely — but the judge-and-refine loop lifts every output toward the same bar. Bring your own intelligence.

04 / OPEN SOURCE

Yours. Open. Local.

One MIT file that lives in your repo and runs in your Claude Code. No subscription, no upload, no vendor lock-in. The studio is yours to keep, fork, and bend.

Inspired by ClaudeDesign · built differently

A product, made into a method you own.

We love what Anthropic shipped. We just wanted it open, runnable on any model, and living in our own terminal. So we rebuilt the mechanism — not the product.

ClaudeDesign ships one curated direction.
→
We explore eight, judge all, merge the best.
Runs on one model, behind a paid subscription.
→
Model-agnostic, MIT, runs in your terminal.
A hosted product you rent.
→
A single file you own, fork, and keep forever.
Quality depends on their best vision model.
→
Quality is a floor built by the panel — on any model.
Same philosophy · open source · model-agnostic · runs anywhere
How it works

Six phases, one shippable page

A Claude Code Dynamic Workflow — deterministic orchestration over many model agents in one JS file. No hand-holding, no single lucky roll.

① BRIEF

Distill a design system

Turn the prompt into shared tokens — fonts, palette, mood — injected into every agent after.

② DIVERGE

Explore many directions

N agents each commit to a distinct aesthetic and ship a complete page. The quality core.

③ JUDGE

Panel scoring

Every direction × 6 dimensions, scored 1–10 by independent critics. Each returns a top fix.

④ SYNTHESIZE

Graft the best

Top-3 directions read and merged into one page that must beat any single direction.

⑤ REFINE

Adversarial polish

Ruthless critique → surgical fix, looped. ClaudeDesign's "fine-grained controls," engineered.

⑥ POLISH

Ship-ready QA

Responsive, A11y, WCAG AA, reduced-motion, no broken refs — verified in headless Chromium.

Live outputs

24 real pages across 24 domains

Genuine pipeline output — no hand-editing. The first 14 are multi-direction runs (winners span Bold Editorial ×8, Dark Premium ×3, Swiss Minimal ×2); the 10 below are single-lens style studies, each aesthetic paired with a domain that fits it.

Full deep-dives →
Multi-direction examples · 14
🎬 MOVIE

LUMIÈRE

Cinematic rating & discovery. Hero 9.2 score, trending, charts, genres.

Dark Premium49 agents
🍽️ DINING

MAISON NOIR

Fine dining. Tasting menu, candle-lit cellar, antique gold.

Dark Premium35 agents
🎨 STUDIO

OBJECT & ECHO

Experimental agency. Kinetic type, object/echo duality, acid signal.

Bold Editorial35 agents
🎵 MUSIC

NOVA · AURORA

Immersive album. Generative cover, 32-bar visualizer, play-state aura.

Bold Editorial35 agents
💪 FITNESS

PULSE

Training app. Adaptive plans, data-minimal athletic aesthetic.

Swiss Minimal35 agents
✈️ TRAVEL

AZORES

Destination guide. Interactive islands map, big editorial photography.

Bold Editorial35 agents
📚 EDTECH

LUMEN

Learning platform. Course cards, learning-path journey animation.

Swiss Minimal35 agents
🎮 GAME

ECHOES OF THE VOID

AAA promo. Ship-status HUD, cinematic title reveal, editions.

Bold Editorial35 agents
🤝 NONPROFIT

Brightwater

Clean-water cause. Animated donation well, transparency donut.

Editorial35 agents
🛒 COMMERCE

AURA ONE

Premium headphones. Variant picker, instrument-grade specs.

Dark Premium35 agents
🎤 CONFERENCE

BUILD/2026

Dev conference. Multi-track schedule, build-pipeline graph.

Bold Editorial29 agents
🗄️ SAAS

vector

Vector DB landing. Benchmark stats, Python snippet, pricing.

Bold Editorial49 agents
⌨️ OSS

tideline

CLI homepage. Install command, terminal demo, monospace.

Bold Editorial35 agents
🖼️ PORTFOLIO

Lin Hua

Designer portfolio. Stat hero, 6-project grid, testimonials.

Bold Editorialenriched
Style study — 10 styles × 10 themes · 10
01 / SYNTHWAVE

PIXEL DIVE

Retro arcade & bar — cabinets, game-named cocktails, tournaments.
02 / BAUHAUS

FORM & FUNCTION

Design academy — courses, exhibition, manifesto.
03 / ART DECO

THE GILDED HOUR

1920s speakeasy — cocktails, jazz, reservations.
04 / MEMPHIS

RIOT DECK

Party card game — how to play, expansions, reviews.
05 / RISOGRAPH

INK & PAPER

Indie zine & print studio — issues, print shop, subscribe.
06 / SCANDI

HYGGE & CO

Nordic home goods — products, story, materials.
07 / HOLOGRAPHIC

PRISM

Holographic cosmetics — products, shades, looks.
08 / COUTURE

VESPER

Couture A/W — lookbook, atelier, private viewing.
09 / PUNK ZINE

BLACKOUT

Punk band — tour dates, new record, merch.
10 / CLAYMORPHISM

BLOOM

Kids' learning app — how it works, subjects, for parents.
FAQ

Questions, answered.

Do I need a specific model?
No. Every agent inherits your current session model. Stronger models explore more sharply, but the judge-and-refine loop lifts any model toward the same bar — that's the whole point.
How long does a run take?
Roughly 15–30 minutes for 4–6 directions. It's deliberate: ~30–50 agents explore, judge, synthesize, and polish in parallel. Quality is the point, not latency.
What do I actually get?
One self-contained final.html (single file, no build), the explored draft-*.html directions, and a short summary of which aesthetic won and why.
Can I use just one direction?
Then you don't need this workflow — run a normal frontend pass. The value is the parallel exploration + independent judging; remove that and it's just a generator.
Can I plug in my own design system?
Yes. Pass brand notes (colors, fonts, references) via the brand argument; the brief phase folds your tokens into every agent so the whole pipeline stays on-brand.
Is it really free?
MIT, open source, runs locally in your Claude Code. If it helps you ship, consider sponsoring — it funds new lenses, more examples, and ongoing polish.

Stop betting on one roll.

Install the workflow, point it at any brief, and let a panel of design critics do the exploration you never have time for.

Star on GitHub Sponsor the project Read the deep-dives