AGENT SKILL

ai-design-generation

Produce visual assets from AI image and UI models through MCP — Stitch screens, Fal.ai (Imagen, FLUX, Veo), Recraft V3 vectors, GPT-Image and Figma file access — with prompt patterns and quality scoring. Use when rendering mockups, icons, illustrations or video. Not for interface copy; not for agent UX.

License 许可
MIT
最近更新
2026-11-10
Stars
★ 0
Forks
0

01 前言 Front Matter

name
ai-design-generation
description
Produce visual assets from AI image and UI models through MCP — Stitch screens, Fal.ai (Imagen, FLUX, Veo), Recraft V3 vectors, GPT-Image and Figma file access — with prompt patterns and quality scoring. Use when rendering mockups, icons, illustrations or video. Not for interface copy; not for agent UX.
allowed-tools
Read, Grep, Glob

02 资源文件 5

03 SKILL.md 正文

AI Design Generation

Purpose

Turn a design brief into AI-generated screens, icons, illustrations, photos and video, then curate them to production quality. The tools have no understanding of users, goals or brand; the agent supplies the brief, constraints, quality bar and iteration direction. MCP servers are optional: without them, produce detailed specs and prompts the user can run in any tool.

Core rules

  • Run the curation loop. Specify → generate 3-6 candidates → evaluate against brand, usability and aesthetics → select → refine the prompt → repeat. 2-4 rounds usually reach production quality. Never ship a first-generation output unreviewed.
  • Write prompts like briefs. Be specific ("16px Inter Medium, #1A1A2E on white, 8px radius card", not "modern and clean"). Name real products ("Stripe dashboard density"). State what the design must NOT do. Layer the prompt: style, layout, content, details. Inject design tokens, sector and platform (for iOS 26 Liquid Glass or Material 3 Expressive conventions, see platform-visual-standards).
  • Pick the tool by asset: UI screens → Stitch; SVG icons and vectors → Recraft V3; hero photos → FLUX Pro or Imagen 4; editorial illustration → Midjourney (manual) or FLUX; concepts and patterns → FLUX Schnell; video → Veo 3.1 via fal. OpenAI image models have no official MCP server, so call the Images API or hand the user the prompt.
  • Go from breadth to depth. Use cheap, fast models at low resolution for 10-20 concepts, premium models for 4-6 variations of the chosen directions, then refine, and upscale only the selected outputs.
  • Keep sets consistent. Generate an anchor (the most complex icon or the hero illustration) first, batch the rest against it in groups of about 6, review the set as a whole, and regenerate outliers.
  • Quality gate. Check for garbled text (replace it with real copy), spacing drift, phantom UI, style drift, anatomical errors, brand contamination and soft edges. Score brand compliance on 7 dimensions (0-10). Production needs a 7+ average. Color-correct to brand values and re-check WCAG contrast.
  • Hand off. Rebuild generated screens with Figma Auto Layout and real tokens, or map them to existing components and code via component-patterns-code, then compare the rendered output against the mockup.
  • Ethics. Prompt explicitly for demographic and cultural diversity. Audit accessibility, since generated designs rarely meet WCAG. Verify commercial-use licenses. Tell clients when shipped or presented work is AI-generated.

Routing

Question Read
Which MCP server; configuration; Stitch, fal, Recraft, Midjourney, DALL-E, Veo, Runway, Pika capabilities and prompt patterns references/mcp-design-tools.md
Prompt anatomy per tool, token injection, sector/platform prompting, negative prompts, UI/icon/illustration/landing-page templates references/prompt-engineering-visual.md
Multi-step and batch pipelines, design-system asset generation, Figma import, code from mockups, when to hand-craft references/generation-pipelines.md
Scoring output, artifact detection, consistency checklist, brand-compliance rubric, iteration fixes references/quality-control-generation.md
Icons, illustrations, photos, video, patterns; upscaling, background removal, SVG tracing, color correction; ethics; cost per model and asset references/asset-generation-guide.md

Cross-References

  • visual-design-mastery — Provides the design intelligence that makes generation prompts effective and quality scoring meaningful.
  • platform-visual-standards — Provides platform-specific constraints for generation prompts.
  • ui-pattern-intelligence — Identifies which patterns to generate and their quality benchmarks.
  • sector-style-intelligence — Provides sector-specific visual direction for generation prompts.
  • design-systems-architecture — Ensures generated designs consume design tokens consistently.
  • figma-design-tool-workflows — Integration with Figma for design-to-code pipelines.
  • icon-illustration-systems — Detailed icon and illustration design guidance that informs generation prompts.
  • color-palette-library — Curated palettes for color-accurate generation prompts.
  • typography-pairing-recipes — Type system knowledge for typography-aware generation.
  • animation-recipe-library — Motion design knowledge for video generation prompts.
  • scroll-experience-direction — Kie AI stills, image-to-video shots and background removal for scroll sites (kie-ai.md).

Commands Powered by This Skill

Command How This Skill Is Used
/generate Primary command — screen and asset generation (icons, illustrations, photos, video) via MCP tools
/grade Can generate mood board visualizations using image generation
/screen Can use Stitch to generate screen mockups before coding
/component Can reference generated designs as implementation targets

04 同仓库更多 Skills