One-of-a-Kind Design

Instructions

Step 0: Setup (first run only)

CRITICAL: On first invocation, install dependencies before anything else:

bun run .claude/skills/one-of-a-kind-design/scripts/setup.ts

This installs all runtime dependencies (Mastra, fal.ai, E2B, Effect, pixelmatch, etc.) and creates a

.env

template. The user must fill in API keys before generation will work.

Skip this step if

node_modules/@mastra/core

already exists.

Step 1: Parse the Request

The user invokes the skill with

/one-of-a-kind-design

followed by their request.

Interactive mode (default):

/one-of-a-kind-design Design a site for my omakase restaurant

Headless mode (no human input, for CI/CD or batch processing):

/one-of-a-kind-design --print Design a site for my omakase restaurant

The Mastra workflow automatically:

• Extracts output type, industry, mood, audience from the intent
• Computes specificity (0-7). If below 3, ask the user 1-3 clarifying questions
• Resolves style from taxonomy with convention-breaking detection

Expected output: streaming progress events for each pipeline step.

Step 2: Review Pipeline Output

The workflow streams progress for each step:

• resolve-style -- Maps intent to style config from 65+ styles in the taxonomy
• select-models -- Picks optimal fal.ai endpoints based on style affinity
• craft-prompt -- Agent-powered prompt crafting with style tokens, palette hex codes, composition directives
• generate-artifact -- fal.ai generation (Flux Pro, Seedance 2.0, Kling, Recraft, etc.)
• post-process -- E2B sandbox processing (sharp, potrace, SVGO optimization)
• verify -- 4-layer verification (pixelmatch, SSIM, pHash, uniqueness)
• score-quality -- LLaVA 13B vision scoring, 9 weighted sub-scores, composite 0-10

Hero asset archetypes are auto-selected based on style motion_signature:

Archetype	When	Pipeline
Panning Scene	cinematic, editorial	Video camera choreography
Parallax Depth Stack	layered, atmospheric	Image generation plus depth estimation
Generative Canvas	algorithmic styles	Image generation
3D Object Showcase	product, isometric	3D mesh generation
Typographic Statement	editorial, swiss	Image generation
Photographic Drama	double-exposure, infrared	Image generation
SVG Vector Graphic	logos, icons, decorative	QuiverAI Arrow vectorization

Step 3: Quality Gate

Composite score is computed from 9 weighted sub-scores:

Sub-score	Weight
Anti-slop gate	0.15
Prompt-artifact alignment	0.15
Aesthetic	0.13
Style fidelity	0.13
Distinctiveness	0.13
Asset quality	0.12
Hierarchy	0.06
Convention break adherence	0.05
Color harmony	0.05
Code standards	0.03

Minimum: 7.0/10. If below, the workflow suspends in interactive mode. Present the scores to the user and ask: accept, retry with feedback, or adjust dials. In headless mode, auto-retries up to 3 times with seed bumps.

Step 4: Deliver

Share the generated artifact with the user. The artifact includes:

• Generated hero asset (image, video, SVG, or 3D mesh)
• Full website/app code using resolved style's Tailwind v4 preset
• Quality score card with all 9 sub-scores
• Style metadata (ID, palette, motion signature, premium patterns)

Headless / CI Mode

For automated pipelines without human interaction, use the

--print

flag:

/one-of-a-kind-design --print Album cover for a jazz trio

In

--print

mode:

• No suspend, no quality gate prompts, no human-in-the-loop
• Auto-retries up to 3 times with seed bumps if quality is below 7.0
• Outputs the artifact URL and composite score when done
• Exit code 0 for success, 1 for failure

For CI/CD scripts:

bun run .claude/skills/one-of-a-kind-design/scripts/mastra/modes/ci.ts '{"userIntent":"...","outputType":"..."}'

MCP Integration

Tools are exposed via MCP for any MCP-compatible client:

• design-generate

  -- Full pipeline execution

• design-resolve-style

  -- Style resolution only

• design-score

  -- Quality scoring only

• design-verify

  -- Verification only

Anti-Slop Rules

Full reference:

references/ANTI-SLOP.md

. Top rules:

1. NO Inter/Roboto/Open Sans -- use style-specific typography
2. NO purple-to-blue gradients -- derive palette from style
3. NO default shadows -- use style's shadow model
4. NO hero/features/testimonials/CTA skeleton -- restructure per style
5. NO linear easing -- spring physics or style curve
6. NO #000000 body text -- off-black (#111, #1a1a1a, #2F3437)
7. NO Lorem Ipsum/Acme Corp -- realistic content
8. NO "Elevate/Seamless/Unleash" -- specific, human copy
9. NO identical card grids -- vary sizes for hierarchy
10. NO simultaneous element mount -- stagger entry (80ms delay)

Environment

Requires environment variables:

FAL_KEY

,

E2B_API_KEY

,

QUIVERAI_API_KEY

.

All scripts use Bun-only APIs, Effect-native TypeScript, max nesting depth 3.

Examples

Example 1: Restaurant Website

User says: "/one-of-a-kind-design Design a website for my omakase restaurant in Brooklyn. No food photos."

Actions:

1. Resolve style: wabi-sabi (food + warm + intimate), variance 7, convention break: no food photography
2. Select model: Flux Pro 1.1 (cinematic affinity), chain: t2i
3. Craft prompt: 280 chars, subject-first, hex palette #8B7355 #D4C5A9 #F5F0E8
4. Generate hero image via fal.ai, seed 42, E2B converts to WebP
5. Verify: pHash stored, uniqueness confirmed (hamming 42 to nearest)
6. Score: LLaVA 13B composite 7.9/10 PASS

Result: Full website with hand-textured hero, no food photography, seasonal palette system.

Example 2: Album Cover

User says: "/one-of-a-kind-design Album cover for a jazz trio's debut. Smoky, intimate, blue."

Actions:

1. Resolve style: cinematic (jazz + intimate via compound map)
2. Select model: Flux Pro 1.1, chain: t2i
3. Craft prompt: atmospheric lighting, rich shadows, palette #1A1A2E #0F3460 #E94560
4. Generate, E2B post-process, verify uniqueness
5. Score: LLaVA 13B composite 7.7/10 PASS

Result: Cinematic album artwork with atmospheric depth and jazz-appropriate color theory.

Example 3: Video Trailer

User says: "/one-of-a-kind-design 15-second trailer for a contemporary opera staging deconstructed Madama Butterfly"

Actions:

1. Resolve style: deconstructivism, chain: t2i-i2v (keyframe then animate)
2. Generate keyframe via Flux Pro, animate via Seedance 2.0 (duration "15", aspect "21:9")
3. E2B extracts first frame for quality verification
4. Score: LLaVA 13B composite 7.2/10 PASS

Result: 15s cinematic video with fractured visual language and camera choreography mirroring emotional arc.

Example 4: SVG Logo

User says: "/one-of-a-kind-design Logo for a sustainable fashion brand called Thread"

Actions:

1. Resolve style: editorial-minimalism (fashion + minimal)
2. Route to SVG pipeline: Recraft V3 vector illustration mode
3. E2B runs SVGO optimization, tests at 16px and 4K
4. Score: LLaVA 13B composite 7.5/10 PASS

Result: Single-color vector logo with real SVG paths, tested across sizes.

Example 5: Mobile App

User says: "/one-of-a-kind-design Onboarding screens for a language learning app targeting adults over 40"

Actions:

1. Resolve style: material-design (education + clean)
2. Generate hero texture, E2B creates phone-frame mockup
3. Verify uniqueness, score quality
4. Score: LLaVA 13B composite 7.4/10 PASS

Result: Accessible onboarding screens with warm typography and generous touch targets.

Example 6: Headless / CI

User says: "/one-of-a-kind-design --print Event poster for a warehouse techno party in Berlin"

Actions:

1. Headless mode: no suspend, no human input
2. Resolve style: glitch (techno + underground), generate via Flux Pro
3. Auto-score: composite 6.8 (below 7.0), auto-retry with seed 43
4. Re-score: composite 7.3/10 PASS on second attempt

Result: Poster with neon typography, scan lines, RGB split. Delivered with score and URL, exit code 0.

Troubleshooting

Error: fal.ai returns 404 Cause: Model endpoint deprecated Solution: Run

bun run scripts/mastra/tools/index.test.ts

to verify working endpoints

Error: Quality score below 7.0 Cause: Prompt does not match the resolved style well enough Solution: In interactive mode, provide feedback. In headless mode, auto-retries 3 times with different seeds.

Error: Pixelmatch shows 0% diff after refinement Cause: Generator did not apply feedback Solution: Pipeline auto-escalates to higher-tier model with explicit fix instructions

Error: Alignment below 5.0 after 3 retries Cause: Prompt-model mismatch or overly abstract prompt Solution: Escalate to higher-tier model. If still failing, flag for manual review.

Error: Style conflict detected Cause: Two incompatible styles combined Solution: Check

references/CONFLICT-MAP.md

for resolution patterns.

Performance Notes

• Take your time with style resolution. Read the full taxonomy entry before resolving.
• Do not skip validation steps even when confident. Every generation gets alignment-checked.
• Quality over speed. A 7.0 composite reached in 3 attempts beats an 8.0 achieved by luck.
• Fix one quality dimension at a time.
• After 2 failed regenerations at the same tier, escalate exactly one tier.
• Maximum 15 fal.ai calls per pipeline run.
• Never fabricate alignment scores or quality sub-scores.

Video Generation Reference

For video output types, consult these reference documents:

• references/VIDEO-SCENE-SCHEMA.md

  — Complete scene ontology (entities, composition, motion, lighting, materials, atmosphere, audio, narrative, editing), JSON/YAML schemas, prompt templates for Runway/Sora/Pika styles, identity locks, continuity locks, motion realism blocks, anti-failure blocks, and evaluation rubric.

• references/VIDEO-ELEMENT-TEMPLATES.yaml

  — Reusable element templates with shared enums (realism levels, framing layers, camera angles, lens types, motion types, materials, surface finishes), validation primitives, reference image roles, and clip assembly templates for multi-shot continuity.

When generating video:

1. Build a scene schema using the ontology (entities, composition, lighting, motion)
2. Use the appropriate prompt template variant (Runway-style, Sora-style, etc.) based on the selected model
3. Apply identity lock and continuity lock blocks for multi-shot consistency
4. Include the anti-failure block in negative prompts
5. Evaluate output against the rubric in section 9

Surgical Image Editing (Nano Banana Pro Edit)

When brief compliance check finds missing elements, or the user requests a specific edit to a generated image, use Nano Banana Pro Edit for surgical, localized changes that preserve the existing composition.

bun run scripts/edit-with-nano-banana.ts --image <url> --edit "Add three nail holes in the plaster surface, catching amber side-light"

The edit pipeline:

1. Enhances the edit description into a precise Nano Banana prompt (enforcing spatial locality, dimensional integrity, photometric consistency)
2. Executes the edit via

   fal-ai/nano-banana-pro/edit

   with

   image_urls

   array
3. Returns the edited image URL

Use this INSTEAD of image-to-image for adding missing elements. Nano Banana preserves the base image exactly — i2i re-generates too much.

Endpoint:

fal-ai/nano-banana-pro/edit

Input:

{ prompt: string, image_urls: string[], seed: number }