chenxin-corelisted
Use this skill whenever a ComfyUI / generative-model request comes in for Claude Code. Trigger on ANY of: "comfyui", "comfy ui", "workflow", "出图", "跑工作流", "生成图片", "生成视频", "出视频", "manga", "漫画", "anime", "anima", "wan", "ltx", "ltx-2.3", "hunyuan", "flux", "sdxl", "sd 1.5", "sd1.5", "stable diffusion", "krea", "seedream", "nano banana", "Qwen-Image", "ideogram", "Recraft", "Kling", "Seedance", "Veo", "Sora", "Runway", "Luma", "Stable Audio", "ACE-Step", "video", "talking head", "inpaint", "upscale", "controlnet", "IP-Adapter", "refiner", "LoRA", "8 GB VRAM", "8GB", "small VRAM", "low VRAM", "vae", "unload model". This is the L4 mega-skill — it owns the dispatch table from a single Claude Code entry point to L1 (ComfyUI runtime), L2 (MCP driver), L3 (knowledge substrate: 80 recipes + 662 templates + 8 GB hardware matrix), and L5 (application skills, future P1.1). For ANY generative-model prompt, read this skill first to find the right tool, recipe, and VRAM-safe defaults before writing the prompt or invoking a w
cxin21/comfyui-chenxin · ★ 0 · AI & Automation · score 70
Install: claude install-skill cxin21/comfyui-chenxin
# chenxin-core — L4 mega-skill
This skill is the **single entry point** for any ComfyUI / generative-model
work in Claude Code. It owns the routing from a user request to the right
combination of L1 / L2 / L3 / L5 tool, recipe, and hardware defaults.
It exists because P0.1 + P0.2 gave us the pieces (recipes, templates,
hardware matrix, MCP tools) but no single place to dispatch from. Without L4
the rest is dormant content.
## What L4 does
1. **Routes keywords → tools.** "出动漫角色" is NOT a single command —
it's a recipe lookup + a VRAM decision + a workflow selection + a
generation call. L4 owns this composition.
2. **Enforces the auto-pull rule.** When a specific model is named in the
request, L4 reads its recipe dialect (via `internals/recipe_lookup.py`)
BEFORE the prompt is written — never after.
3. **Overlays recipe overrides on hardware defaults.** `internals/hardware_decide.py`
layers recipe-level `quant` / `steps` / `cfg` / `scheduler` on top of the
P0.2 hardware matrix.
4. **Prefer Chinese-first naming, English-fallback search.** When the user
writes "动漫" or "漫画", prefer Chinese-canon model families first
(Anima, Wan, LTX-2.3, Hunyuan); fall back to Western canon (FLUX, SDXL)
only if the user signals it explicitly.
## When to invoke this skill
Invoke chenxin-core **before** doing ANY of the following:
- Calling `mcp__comfyui-mcp__*` tools (L2)
- Writing a generation prompt for a specific named model
- Selecting a workflow template (L3 te