What is the ai-design-generation skill?
The ai-design-generation skill is a SKILL.md in phazurlabs/sumi for producing visual assets from AI image and UI models. Your AI coding agent gets prompt patterns for Stitch, Fal.ai, Recraft V3, Midjourney and DALL-E 3, a curation loop, and a 0 to 10 scoring scale.
The ai-design-generation skill produces visual assets from AI image and UI models through MCP: Stitch screens, Fal.ai with Imagen, FLUX and Veo, Recraft V3 vectors, GPT-Image and Figma file access. It adds prompt patterns and quality scoring.
It works as a human curation loop: specify, generate 3 to 6 candidates per round, evaluate, select, refine and repeat. Two to four rounds typically reach production quality, with prompting treated as a design skill.
MCP servers and models
- The supported MCP servers, all optional, are stitch-mcp, @fal-ai/mcp, replicate-mcp, openai-images-mcp and figma-mcp, and each needs an API key.
- Fal.ai gives access to 600+ models. FLUX Pro reaches 2048x2048, with steps of 28 to 50 for Pro and 4 for Schnell, and guidance of 3.5 to 7.5.
- Recraft V3 produces true SVG vector output, with four illustration presets including corporate flat and isometric.
- Without MCP servers the skill provides detailed specifications and prompts to use manually with any generation tool.
Pipelines and quality control
Pipelines run in four stages: concept exploration, variation, refinement and production. An icon set pipeline exports SVG at 16, 20, 24, 32 and 48 grid sizes, and brand compliance scoring on a 0 to 10 scale needs a 7 or higher average to ship.
Who the ai-design-generation skill is for
Best for
- Rendering UI mockups, icons, illustrations or video from AI image and UI models.
- Teams that want a repeatable prompt structure layering style, layout, content and details.
- Checking generated assets for text garbling, style drift and brand contamination before use.
Not for
- Building a multi-page site from Google Stitch by reading a baton file and writing the next baton, which the stitch-loop skill covers.
- Writing a DESIGN.md that describes a Google Stitch project's look in plain language, which the design-md skill covers.
- Writing a .stitch/SITE.md memory file with identity, sitemap and roadmap for the Stitch build loop, which the site-md skill covers.
How to install AI Design Generation Skill: Prompts and Quality Scoring
/plugin marketplace add phazurlabs/sumi
/plugin install sumi@sumi-marketplacegit clone https://github.com/phazurlabs/sumi.git && claude --plugin-dir ./sumi.Check the requirement
You need the Claude Code CLI, authenticated. The core is markdown with zero config, and it works offline.
Add the marketplace and install
Run the two commands above to add the phazurlabs/sumi marketplace and install the sumi plugin.
Pick a scope
Choose the scope when you install: user for all your projects, project shared through .claude/settings.json, or local, which is yours and gitignored.
Start
Run
/sumi:start, which asks what you are working on in plain words and picks the play. For image generation, add any of the optional MCP servers and set its API key.
Check that it loaded
After installing, run /sumi:start in your agent. If /plugin is not recognised, update Claude Code.
Key features
- Human curation loopSpecify, generate 3 to 6 candidates per round, evaluate, select, refine and repeat. Two to four rounds typically reach production quality.
- Prompt principlesSpecificity beats length, reference real products, constrain before creating, and layer style, then layout, then content, then details into one coherent prompt.
- Optional MCP serversstitch-mcp, @fal-ai/mcp, replicate-mcp, openai-images-mcp and figma-mcp are supported, each with its own API key, and none is required.
- Fal.ai settingsFLUX Pro supports up to 2048x2048 with steps of 28 to 50, FLUX Schnell uses 4 steps, and guidance runs from 3.5 to 7.5, where higher means more prompt adherence.
- Recraft V3 vectorsTrue SVG vector output, an icon prompt structure and four illustration style presets, including corporate flat, plus a style anchor technique.
- Midjourney and DALL-E 3 parametersMidjourney prompts use --ar, --style raw, --stylize and --no. DALL-E 3 resolution is limited to 1024x1024, 1024x1792 or 1792x1024.
- Video modelsVeo 3.1 via Fal.ai generates clips for prototype animations, product demos and marketing content, alongside Runway Gen-3 Alpha and Pika.
- Four stage pipelinesConcept exploration, variation, refinement and production, with an icon set pipeline that exports SVG at 16, 20, 24, 32 and 48 grid sizes.
- Quality controlSeven artifact types to catch, an eight item cross-screen consistency checklist, and brand compliance scoring from 0 to 10 across seven dimensions with a minimum average of 7.
- Post-processing and costUpscaling with Real-ESRGAN, background removal with remove.bg, SVG tracing with Vectorizer.ai, and a cost per asset table, such as $0.05 typical for a UI screen mockup.
Frequently asked questions
What is the ai-design-generation skill?
It is a SKILL.md in the phazurlabs/sumi repository for producing visual assets from AI image and UI models through MCP. Your AI coding agent gets prompt patterns, a curation loop and quality scoring for mockups, icons, illustrations and video, with 3 to 6 candidates generated per round.
How do I install ai-design-generation?
Run /plugin marketplace add phazurlabs/sumi, then /plugin install sumi@sumi-marketplace, and start with /sumi:start. Pick the scope when you install: user for all your projects, project through .claude/settings.json, or local and gitignored. The core is markdown with zero config. If /plugin is not recognised, update your agent.
Do I need MCP servers to use it?
No. The supported MCP servers are all optional: stitch-mcp, @fal-ai/mcp, replicate-mcp, openai-images-mcp and figma-mcp, each needing its own API key. Without MCP servers, the skill provides detailed specifications and prompts that you can use manually with any generation tool. Fal.ai alone gives access to 600+ models.
Which generation tools and models does it cover?
It covers Stitch for UI screens, Fal.ai with Imagen 4, FLUX and Veo 3.1, Recraft V3 for true SVG vectors, Midjourney, DALL-E 3, Runway Gen-3 Alpha and Pika. Each has prompt guidance, such as FLUX Pro up to 2048x2048 and DALL-E 3 resolutions of 1024x1024, 1024x1792 or 1792x1024.
How does it score generated designs?
Generated designs are scored against brand requirements on a 0 to 10 scale across seven dimensions. The minimum viable score for production use is an average of 7 or higher across all dimensions. It also lists seven artifact types to check, including text garbling, style drift and brand contamination.
Which agent do I need and does it work offline?
You need the Claude Code CLI, authenticated. The core of the plugin is markdown with zero config, and it works offline. Run /sumi:start after installing and it asks what you are working on in plain words, then picks the play, so you never have to learn skill names.