The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill requires the installation of the @runcomfy/cli Node.js package from the NPM registry to function. It also facilitates the download of generated video and audio assets from vendor-controlled domains including *.runcomfy.net and *.runcomfy.com.
[COMMAND_EXECUTION]: The skill relies on executing the runcomfy command-line tool via the shell to interact with the ByteDance Seedance model API. User-provided prompts and configuration are passed as arguments to this CLI.
[INDIRECT_PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection as it ingests untrusted data from users that is then processed by a generative model.
Ingestion points: Untrusted data enters the agent context through the prompt, image_url, video_url, and audio_url fields defined in SKILL.md.
Boundary markers: The instructions mandate wrapping user prompts inside a JSON object (--input '{"prompt": "<user prompt>"}') before passing them to the CLI, which provides a structural boundary.
Capability inventory: The skill has the capability to execute shell commands (runcomfy run ...) and download files to a local --output-dir.
Sanitization: The skill documentation claims the CLI avoids shell expansion by transmitting the JSON body directly over HTTPS, although it does not implement specific sanitization for adversarial instructions embedded within the prompt text or multi-modal reference files.

seedance-v2