The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes the belt CLI tool within a Bash environment to perform AI inference tasks. This includes commands for authentication (belt login) and executing specific AI models (belt app run) to generate media.
[EXTERNAL_DOWNLOADS]: The skill references documentation and asset samples from external domains including inference.sh and GitHub. It also provides instructions to extend environment capabilities using the npx skills add command for the belt-sh/cli and inference-sh/skills repositories.
[INDIRECT_PROMPT_INJECTION]: The skill functions as a pipeline for external media and text inputs, which introduces a potential surface for indirect prompt injection targeting the remote AI models.
Ingestion points: User-provided image URLs, audio URLs, and text scripts (e.g., voice_script, voice_prompt) passed to the belt CLI.
Boundary markers: Inputs are structured as JSON parameters, but the natural language fields lack explicit delimiters to separate user-provided content from system instructions in the underlying model prompts.
Capability inventory: The skill uses the belt CLI tool to interact with remote services; the tool's access is scoped to the belt command via the allowed-tools configuration.
Sanitization: No specific input sanitization or validation logic is defined within the provided skill instructions.

ai-avatar-video