The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the user to download model weights from a third-party Hugging Face mirror (huggingface.co/Aitrepreneur/FLX) instead of the official Baidu or Comfy-Org repositories. The skill author explicitly notes that the provided installer scripts (.bat and .sh) download extra files for unrelated models (Z-Image).
[REMOTE_CODE_EXECUTION]: The installation process involves cloning multiple third-party GitHub repositories to add custom functionality to the ComfyUI environment. These repositories (e.g., ltdrdata/ComfyUI-Manager, city96/ComfyUI-GGUF, rgthree/rgthree-comfy, etc.) are managed by individual community members and are not within the platform's trusted organization list. Additionally, the skill refers to automated installer scripts that execute shell commands to manage files and dependencies.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data in the form of user prompts and ComfyUI workflow JSON files. These inputs influence the behavior of the image generation pipeline and custom nodes.
Ingestion points: User-supplied text prompts and .json workflow files specified in the globs field.
Boundary markers: None identified; instructions do not include specific delimiters or warnings to ignore instructions embedded in the prompt data.
Capability inventory: The skill environment includes file system writes (storing models/nodes), network requests (fetching external assets), and execution of Python-based custom nodes.
Sanitization: There is no evidence of validation or sanitization for the structure of the ingested JSON workflows or the content of the text prompts.

ernie-image