The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides instructions for the agent to invoke the runcomfy command-line utility. It uses a structured approach, passing user inputs such as prompts and URLs within a JSON object to the CLI's --input flag. The documentation notes that this method is intended to prevent shell injection by avoiding direct shell expansion of the prompt content.
[EXTERNAL_DOWNLOADS]: The skill documentation requires the installation of the @runcomfy/cli package from the public NPM registry. It also describes the CLI's behavior of downloading generated video files from vendor-specific domains (runcomfy.net and runcomfy.com) to a user-specified output directory. These activities are consistent with the skill's stated purpose of interacting with the RunComfy service.
[INDIRECT_PROMPT_INJECTION]: As the skill is designed to process untrusted user-supplied data—including text prompts and external media URLs—it presents a surface for indirect prompt injection attacks at the model layer.
Ingestion points: User-provided inputs are accepted for image_url, video_url, audio_url, and prompt fields as detailed in SKILL.md routes.
Boundary markers: The skill uses JSON encapsulation as a boundary between the user-supplied string and the shell command execution.
Capability inventory: The skill has capabilities for shell command execution (runcomfy), network access (handled by the CLI), and file system writes (via the --output-dir parameter).
Sanitization: The instructions rely on JSON stringification provided by the agent's execution environment to escape potentially malicious characters before they are passed to the CLI.

image-to-video