The Agent Skills Directory

[COMMAND_EXECUTION]: The skill interacts with the belt CLI tool to perform model inference and application management. Usage is restricted to the belt command via the allowed-tools metadata, which follows the principle of least privilege by limiting the potential attack surface for arbitrary command execution.
[EXTERNAL_DOWNLOADS]: The skill facilitates the installation of the belt-sh/cli and other related skills using the npx skills add command. It also fetches configuration and documentation from the official inference-sh GitHub organization and uses remote assets for its documentation.
[INDIRECT_PROMPT_INJECTION]: The skill handles user-supplied prompts and media URLs by interpolating them into JSON structures for CLI execution. While this represents an attack surface for indirect instructions, the risk is mitigated by the structured format and tool-specific execution.
Ingestion points: User inputs for prompts, image URLs, and audio URLs are passed to the belt app run command via the --input flag.
Boundary markers: JSON formatting is used to provide structure and delimit user data from command arguments.
Capability inventory: The skill has the capability to execute the belt command (restricted bash) to interact with AI inference services.
Sanitization: The skill relies on the agent platform's standard shell escaping and parameter handling for command execution.

p-video