The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied text through various Text-to-Speech (TTS) models, which is a common vulnerability surface for indirect injection.
Ingestion points: The prompt and text fields within the JSON payload for belt app run commands in SKILL.md.
Boundary markers: The skill uses JSON formatting for tool input, providing structural delimitation, though no explicit instructions are provided to the agent to ignore potentially malicious content within the user-provided text.
Capability inventory: The skill uses the Bash tool to execute belt commands for voice generation and media merging.
Sanitization: No explicit sanitization or filtering of the input text is documented.
[COMMAND_EXECUTION]: The skill relies on the execution of the belt CLI tool via the system shell to interact with the inference.sh platform. The allowed-tools configuration in the frontmatter correctly restricts the Bash tool to only execute belt commands, adhering to the principle of least privilege.

ai-voice-cloning