AI-assisted design workflows: prompt engineering for image generation (Midjourney, DALL-E 3, Stable Diffusion, Flux), achieving style consistency across a generated asset set, post-processing AI outputs for production use, legal and licensing considerations, and when AI generation is and isn't appropriate. For teams integrating generative AI into their design workflow.
From clarcnpx claudepluginhub marvinrichter/clarc --plugin clarcThis skill uses the workspace's default tool permissions.
Designs and optimizes AI agent action spaces, tool definitions, observation formats, error recovery, and context for higher task completion rates.
Enables AI agents to execute x402 payments with per-task budgets, spending controls, and non-custodial wallets via MCP tools. Use when agents pay for APIs, services, or other agents.
Compares coding agents like Claude Code and Aider on custom YAML-defined codebase tasks using git worktrees, measuring pass rate, cost, time, and consistency.
✓ Mood boarding — fast visual direction exploration
✓ Background textures and patterns — no identity risk
✓ Abstract illustrations — non-representational art
✓ Concept art for presentations — not for product UI
✓ Placeholder images — to be replaced with real photography
✓ Marketing experiments — low-brand-stakes A/B tests
✓ Reference images for briefing a human designer
✗ Brand identity elements (logo, icon system) — inconsistency, identity risk
✗ Illustrations requiring exact style match to existing set
✗ People with specific demographics — hallucination, bias, legal risk
✗ Realistic-looking fake events, products, or endorsements
✗ Images involving recognizable faces without consent
✗ Medical, legal, or safety-critical visual content
✗ Content that will be passed off as photography without disclosure
[Style adjective(s)] [medium/technique] [subject description],
[color palette / mood], [lighting], [camera/perspective],
[technical quality spec], [negative prompt if supported]
Illustration styles:
flat vector illustration
isometric illustration
line art
editorial illustration
hand-drawn illustration
geometric abstract
Photography styles:
product photography
editorial photography
documentary photography
aerial photography
Art direction:
minimalist
bold and graphic
soft and dreamy
high contrast
muted palette
vibrant
Basic:
/imagine flat vector illustration of a developer at a laptop,
indigo and amber palette, minimal style, white background,
geometric shapes, no shadows --v 6.1 --ar 16:9
Style reference (--sref):
--sref [image URL] # style match to reference image
Character reference (--cref):
--cref [image URL] # maintain character appearance
Stylize (--s):
--s 0 # minimal stylization, follows prompt closely
--s 100 # balanced (default)
--s 750 # strong stylization, more creative
Seed for consistency (--seed):
--seed 12345 # reproducible results for series
Negative prompt (--no):
--no text, watermark, logo, signature, border
Strengths: Strong instruction-following, text in images, safety
Weaknesses: Less stylistic control than Midjourney, can refuse edge cases
Best for: Concept art, marketing copy images, illustrations with text
Prompt tips:
- Be explicit about style: "in the style of a flat vector illustration"
- Specify what NOT to include: "no text, no people, no shadows"
- Request revisions by describing the delta: "same composition but with
a darker background and remove the people"
Flux.1 (Schnell / Dev):
Strong prompt adherence, faster than SDXL
Works well with natural language descriptions
No need for comma-separated token lists
Example (Flux):
A clean isometric illustration of a cloud server rack surrounded by
floating data icons. Deep blue and cyan color palette. Minimal flat style.
Technical, precise. White background. No text. No people.
Stable Diffusion XL:
[positive tags], masterpiece, best quality, 8k
Negative: bad quality, blurry, distorted, text, watermark
ControlNet modes:
Canny → preserve edge structure (for consistent composition)
Depth → preserve 3D structure
Pose → preserve character pose (useful for character consistency)
IP-Adapter → style transfer from reference image
The hardest problem in AI design: making 10 illustrations look like they're from the same series.
1. Create first image: /imagine [prompt] --seed 42
2. Note the job ID and seed from the image details
3. All subsequent images use same seed: --seed 42
4. Variations will share compositional DNA
1. Create approved "reference" image
2. Add --sref [URL] to all subsequent prompts
3. The style (colors, shapes, lighting) transfers
4. Subject still controlled by text prompt
Example:
/imagine empty state for a search feature, friendly character,
indigo palette --sref https://[your-reference-image-url] --sw 100
1. Load IP-Adapter extension in ComfyUI or A1111
2. Use reference image at 0.6-0.8 weight
3. Write prompt for new subject
4. Style and color palette transfer while subject changes
Write a "style fingerprint" — a fixed block you append to every prompt:
[Subject-specific prompt]
Style: flat 2D vector illustration, muted indigo and amber color palette,
simple geometric shapes, round corners, minimal detail, no gradients,
white background, no shadows, no text, no borders, editorial illustration style
Keep this block identical for every image in the series.
AI outputs are almost never production-ready. Budget time for cleanup.
Raw AI output (JPG/PNG, 1024×1024)
↓
Remove background (remove.bg, Photoshop, Figma plugin)
↓
Color correction (match to brand palette in Figma/Photoshop)
↓
Crop to final format (1200×628 OG, 1080×1080 Instagram, etc.)
↓
Compress (Squoosh.app, ImageOptim)
↓
Export at correct resolution (2× for retina)
AI images are raster. For scalable SVG:
1. Adobe Illustrator → Image Trace
Best settings: "16 Colors" preset → expand → ungroup → clean up
2. Vector Magic (vectormagic.com)
Better for illustrations with clear shapes
3. Manual redraw in Figma
For simple shapes — faster than tracing, cleaner result
Rule: Don't vectorize complex photorealistic outputs — only works for
flat, simple illustrations.
In Figma:
1. Import AI image
2. Create overlay rectangle with brand color
3. Set blend mode to "Color"
4. Adjust opacity until palette feels cohesive
5. Export as PNG
In CSS (for web use):
img.brand-tinted {
filter: hue-rotate(30deg) saturate(0.8) brightness(1.1);
}
United States: AI-generated images with no human authorship are
not copyrightable. Works with significant human creative input
(prompting + selection + editing) may qualify.
European Union: Under active regulation. Human authorship required
for full protection.
Practical rule: Treat AI-generated assets as unprotected until
your legal counsel advises otherwise.
| Tool | Training data transparency | Commercial use |
|---|---|---|
| Midjourney | Undisclosed | Allowed (paid plans) |
| DALL-E 3 | OpenAI policy | Allowed (all tiers) |
| Stable Diffusion (open) | LAION dataset | Allowed (check model card) |
| Flux.1 Dev | Black Forest Labs | Research/personal only |
| Flux.1 Schnell | Black Forest Labs | Allowed (Apache 2.0) |
| Adobe Firefly | Adobe Stock (licensed) | Commercially safe |
Safest for commercial use: Adobe Firefly (indemnified), DALL-E 3, Flux.1 Schnell
ai-generated/ folder)| Tool | URL | Strengths | Pricing |
|---|---|---|---|
| Midjourney | midjourney.com | Best image quality, style control | $10-$60/mo |
| DALL-E 3 | via ChatGPT / API | Instruction-following, API access | Pay per use |
| Flux.1 | via replicate.com / local | Open weights, high quality | Free/API |
| Adobe Firefly | firefly.adobe.com | Commercial-safe, Photoshop integration | Creative Cloud |
| Ideogram | ideogram.ai | Best text-in-image generation | Free tier |
| Recraft | recraft.ai | Vector output, style library | Free tier |
| Krea | krea.ai | Real-time generation, style transfer | Free tier |
ai-generated/)