The Agent Skills Directory

[SAFE]: The skill implements the principle of least privilege by using the allowed-tools frontmatter field to restrict the Bash tool. The agent is only permitted to execute commands starting with belt, which limits the potential impact of any command injection attempts.
[EXTERNAL_DOWNLOADS]: The skill references the belt-sh/cli package and provides links to configuration files in the inference-sh/skills GitHub repository. These are legitimate resources required for the primary function of the skill and are documented clearly for the user.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied data through the --input parameter in the belt command examples. The ingestion point is the text field within the JSON input in SKILL.md. Boundary markers are established by the JSON structure itself. The capability inventory is strictly limited to the belt tool as defined in the frontmatter. Sanitization relies on the agent's ability to correctly format shell arguments and the underlying tool's processing of the JSON payload.

elevenlabs-tts