The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes real-time audio from callers, transcribes it using Deepgram, and sends the resulting text to a Groq LLM. The prompt templates in reference/voice-prompts.md interpolate business-specific context which, combined with the lack of explicit delimiters for transcribed user text, represents a potential injection surface.
Ingestion points: Incoming audio streams are handled via Twilio WebSockets and processed by the DeepgramSTTProvider (documented in reference/deepgram-setup.md).
Boundary markers: While the prompt templates include comprehensive behavioral guardrails (e.g., "Never share guest room numbers", "Confirm understanding before taking action"), the code templates do not implement explicit delimiters or "ignore embedded instructions" warnings around the transcribed user speech.
Capability inventory: The skill possesses network capabilities for interacting with STT, LLM, and TTS providers, as well as telephony control via Twilio.
Sanitization: Transcribed text is passed to the LLM without additional escaping or sanitization in the voice_pipeline or the VoiceAgent class.
[SAFE]: The skill demonstrates security best practices by recommending environment variable usage for all API keys and providing clear instructions for webhook signature validation and PII redaction in logs. The use of specific sidecar instructions for performance tracking follows standard metadata reporting patterns and does not involve arbitrary command execution.

voice-ai