The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data during speech-to-text (STT) transcription and text-to-speech (TTS) generation, which constitutes a potential injection surface.
Ingestion points: Untrusted audio data is processed via the Ink STT model in SKILL.md; user-provided text is processed via the Sonic TTS model.
Boundary markers: The skill directs agents to handle transcription output verbatim and warns against manual normalization in SKILL.md, though it does not mandate specific delimiters for data/instruction separation in raw API calls.
Capability inventory: The skill uses network operations (HTTPS/WebSocket) to interact with the vendor API and facilitates file writes (e.g., audio output to /tmp/out.wav).
Sanitization: The skill relies on the underlying Cartesia API's own safety guardrails and does not prescribe additional client-side sanitization of transcription results.
[EXTERNAL_DOWNLOADS]: The skill references and fetches content from the vendor's official repositories and specialized documentation domains.
Fetches SDK READMEs, resource maps, and code examples from the Cartesia GitHub organization repositories.
References a bootstrap script and Markdown documentation from the vendor-owned cartesia.sh and docs.cartesia.ai domains.

cartesia-api