The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted metadata from the public Fish Audio voice library, which represents a potential surface for indirect prompt injection if malicious instructions are embedded in voice titles or descriptions.
Ingestion points: Metadata returned from the GET /fish/model endpoint (e.g., items[] containing titles, tags, and descriptions) is processed by the agent in SKILL.md.
Boundary markers: The skill includes a specific "Gotchas" section that explicitly instructs the agent to treat community voice titles and metadata as untrusted data, providing a natural language boundary.
Capability inventory: The skill utilizes curl for network requests to the api.acedata.cloud endpoint and handles audio URLs; it does not perform direct filesystem writes or command execution on local data.
Sanitization: The skill relies on explicit documentation-level warnings to sanitize agent behavior when interacting with external API data.

fish-audio