The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill references installation instructions and additional functionality from the inference-sh GitHub organization, which is a well-known service for AI model inference.
[COMMAND_EXECUTION]: The skill uses the belt CLI tool to run AI models on remote infrastructure via shell commands.
[INDIRECT_PROMPT_INJECTION]: The skill passes user-controlled strings (e.g., video prompts, image URLs, audio URLs) as arguments within a JSON string to the belt app run command. This configuration establishes a surface where malformed user input containing shell metacharacters could potentially attempt to escape the JSON context and execute unintended commands.
Ingestion points: The prompt, image, and audio fields within the --input JSON payload in SKILL.md.
Boundary markers: Input values are enclosed in JSON double quotes and the entire payload is wrapped in single quotes, providing basic structural separation but no inherent defense against adversarial escape sequences.
Capability inventory: The skill is granted access to the Bash tool restricted to the belt command.
Sanitization: There is no explicit evidence of input validation, escaping, or sanitization within the skill instructions; it relies on the executing agent's internal safety measures.

p-video