The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied text prompts and media inputs (images, audio, video) which are passed as arguments to video generation models and text encoders, such as Gemma 3.1 and UMT5 XXL. This creates a surface for indirect prompt injection where malicious instructions embedded in the input could attempt to bypass safety filters or influence model behavior.
Ingestion points: The --prompt, --input, --audio, and --first/--last arguments in comfy-videogen CLI commands.
Boundary markers: The provided command examples do not use specific delimiters or protective instructions to isolate user-provided prompts from the agent's instructions.
Capability inventory: The skill executes local shell commands, performs file system writes to the outputs directory, and interacts with remote video generation APIs.
Sanitization: There are no explicit instructions for the agent to sanitize or validate the content of user-provided prompts before execution.
[EXTERNAL_DOWNLOADS]: The skill instructions involve the download of external AI model weights to the local environment, which is necessary for model inference.
Evidence: The skill notes that SeedVR2 upscaling "auto-downloads its own model files" (DiT/VAE weights) when they are not present.
Evidence: The skill utilizes comfy-model-downloader to fetch local model files for various generation modes to the /mnt/models/comfyui directory.

comfy-videogen