The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-provided URLs for images and audio files to drive video generation. These external assets represent an attack surface where adversarial content (e.g., text hidden in images or audio commands) could attempt to influence the model's output or the agent's behavior.
Ingestion points: Parameters image_url and audio_url passed to the runcomfy run command as described in SKILL.md.
Boundary markers: The skill instructions specify that the CLI uses structured JSON input and does not perform shell expansion on prompt content, which helps prevent command injection but does not fully eliminate model-level injection.
Capability inventory: The skill is configured to use the Bash(runcomfy *) tool for execution.
Sanitization: The SKILL.md file explicitly warns the agent to ingest only user-provided URLs and to suspect assets if generation results diverge significantly from the prompt.
[EXTERNAL_DOWNLOADS]: The skill instructs the user to install the @runcomfy/cli package via the NPM registry. This is a legitimate utility for accessing the RunComfy service described in the skill's metadata.
[CREDENTIALS_UNSAFE]: The documentation identifies the default storage location for authentication tokens at ~/.config/runcomfy/token.json. It provides best-practice advice for secure token management, such as using environment variables (RUNCOMFY_TOKEN) to avoid persistent file storage in automated environments.
[COMMAND_EXECUTION]: The skill limits its execution capabilities to the runcomfy CLI using the allowed-tools frontmatter field, effectively restricting the agent's ability to run arbitrary shell commands.

ai-avatar-video