The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes speech-to-text transcripts which are untrusted user inputs susceptible to indirect prompt injection attacks. * Ingestion points: Transcripts enter the system via onTranscript (JavaScript) or on_transcript (Python) callbacks. * Boundary markers: The skill recommends deterministic validation and user confirmation to verify intents. * Capability inventory: The server logic is designed to generate agent responses and potentially trigger tools based on these transcripts. * Sanitization: Instructions explicitly warn to treat ASR text as untrusted and derive state safely before processing.
[EXTERNAL_DOWNLOADS]: The installation guide references downloading vendor-specific packages and standard libraries from NPM and PyPI registries.
[CREDENTIALS_UNSAFE]: The skill requires an ElevenLabs API key and provides instructions for managing it using environment variables. It also documents an advanced configuration option to disable JWT authentication, accompanied by strong warnings and requirements for alternative network-level security controls.

speech-engine