The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions and code snippets utilize the child_process.spawn method to execute the rec command (part of the Sox utility). This is used to capture raw PCM audio from the system's default microphone, which is the primary functionality of a streaming transcription skill. The arguments passed to the command are static configuration flags.
[EXTERNAL_DOWNLOADS]: The skill requires the installation of the @deepgram/sdk and express Node.js packages, as well as the system-level utility sox. These are standard requirements for the described audio processing and API integration tasks.
[INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion of external data via real-time audio transcription. While this creates a surface for indirect prompt injection (where instructions spoken into the microphone could potentially be interpreted by an agent processing the transcript), the risk is mitigated by the skill's focus on logging and display rather than automated action execution based on the transcript content.
Ingestion points: Audio data is captured from the microphone via rec and sent to the Deepgram WebSocket in SKILL.md and references/streaming-implementation.md.
Boundary markers: None present for the audio stream or the resulting transcript strings.
Capability inventory: Uses child_process.spawn in SKILL.md to run the recording utility.
Sanitization: No sanitization or filtering is applied to the transcribed text before it is output to the console.

deepgram-core-workflow-b