The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external audio data from streams, local files, and remote URLs (e.g., https://example.com/meeting.mp3) for transcription.
Ingestion points: Audio data enters the agent context through audioStream in the streaming transcription logic and via files or URLs in the pre-recorded transcription logic in SKILL.md.
Boundary markers: The provided code snippets do not include explicit boundary markers or instructions for the agent to ignore embedded commands within the transcribed text.
Capability inventory: The skill transcribes audio and explicitly suggests feeding the final results into an LLM for response generation.
Sanitization: No sanitization, escaping, or filtering of the transcribed content is demonstrated before it is passed to downstream processes.
[EXTERNAL_DOWNLOADS]: The instructions include the installation of official libraries from a well-known service provider.
Evidence: The skill recommends installing @deepgram/sdk via npm and deepgram-sdk via pip to interact with Deepgram's transcription services.

deepgram