The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill references installation guidelines, assets, and CLI tools from the inference-sh GitHub repository and inference.sh cloud storage.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-provided inputs such as image URLs, audio URLs, and text scripts, which creates a surface for indirect prompt injection when these are passed to remote inference models.
Ingestion points: Inputs are accepted via image, audio, and voice_script fields in the JSON payloads passed to the belt CLI tool in SKILL.md.
Boundary markers: External data is structured within JSON objects, which helps delimit data from instructions but does not prevent adversarial content within the strings.
Capability inventory: The skill uses the Bash(belt *) tool to execute remote AI models and interact with the inference.sh API.
Sanitization: The skill does not perform local sanitization of URLs or scripts, relying instead on the security posture of the remote inference service.

ai-avatar-video