ComfyUI Workflow Skill — Generate ComfyUI workflows with natural language via Claude Code

[LingyiChen-AI]

I built a Claude Code skill that generates ready-to-use ComfyUI workflow JSON files through natural language conversation.

Repo: https://github.com/twwch/comfyui-workflow-skill

What it does

Instead of manually connecting nodes in ComfyUI, just describe what you want:

"Generate a FLUX txt2img + Wan 2.2 img2vid workflow"

The skill reads node definitions, selects the right template, and outputs a valid JSON you can drag into ComfyUI and run
immediately.

Key features

• 34 built-in templates — SD1.5 / SDXL / SD3 / FLUX / Wan 2.2 / HunyuanVideo / LTXV / Mochi / Cosmos / Stable Audio / Hunyuan3D
  / Stable Cascade
• 360+ node definitions extracted from ComfyUI source code with exact INPUT_TYPES, defaults, and constraints
• Auto model download — workflows include the native models field, so ComfyUI auto-detects missing models and prompts download
  on import
• LLM integration — supports comfyui_LLM_party nodes for script generation,
  character extraction, prompt enhancement
• Node registry split into 42 category files for on-demand loading (no 4000-line file dump)

Demo

Generate workflow with one sentence

Full pipeline: FLUX txt2img → Wan 2.2 img2vid

Output

default.mp4

Install

git clone https://github.com/twwch/comfyui-workflow-skill.git
ln -s $(pwd)/comfyui-workflow-skill ~/.claude/skills/comfyui-workflow

Then in Claude Code, just say what you want. The skill handles the rest.

How it works

1. SKILL.md defines node specs, class_type mappings, and widget value templates
2. references/nodes/ contains 42 category files (loaders, conditioning, video, audio, 3D, etc.)
3. templates/ has 34 validated JSON templates that serve as base for generation
4. Claude reads the relevant template + node specs, customizes parameters, and outputs valid JSON
5. The models field in the JSON triggers ComfyUI's native missing model dialog

Built with Claude Code. Contributions welcome!