The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external data via image_url and audio_url fields. These assets are untrusted and could potentially contain embedded instructions (e.g., text in images or commands in audio) designed to influence the behavior of the generative models.
Ingestion points: Assets are ingested via the --input JSON payload in commands like runcomfy run bytedance/omnihuman/api as seen in SKILL.md.
Boundary markers: The skill uses structured JSON to pass parameters to the CLI, providing a structural boundary, but it relies on the agent to distinguish between user intent and data content.
Capability inventory: The skill possesses the capability to execute shell commands via the runcomfy tool and write files to the local system using the --output-dir flag.
Sanitization: The skill documentation includes a dedicated security section that warns about these risks and advises agents to only ingest URLs explicitly provided by the user.
[COMMAND_EXECUTION]: The skill facilitates the execution of the runcomfy CLI tool. It defines the allowed scope using allowed-tools: Bash(runcomfy *), which limits the agent to the vendor's specific toolset. It also manages sensitive local files like ~/.config/runcomfy/token.json for authentication, which is standard behavior for the associated service.

ai-avatar-video