The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes shell commands via the Python subprocess module to handle speech playback and interaction with the AI model.
Uses afplay (a standard macOS utility) to play temporary MP3 files generated by the TTS service.
Invokes the claude or codex CLI tools to process the conversation logic in the ask_brain function.
[EXTERNAL_DOWNLOADS]: Performs network operations to interact with ElevenLabs API endpoints for speech processing.
Sends audio data to https://api.elevenlabs.io/v1/speech-to-text for transcription.
Sends text data to https://api.elevenlabs.io/v1/text-to-speech for synthesis.
[PROMPT_INJECTION]: Transcribed user speech is directly interpolated into the prompt sent to the AI engine without sanitization or protective boundary markers.
Ingestion points: Audio captured from the microphone is transcribed into text in voice_mode.py.
Boundary markers: Absent; user transcript is concatenated directly with system instructions in the ask_brain function.
Capability inventory: The resulting prompt is passed to subprocess.run to be executed by the claude CLI.
Sanitization: Absent; the transcribed text is used as-is, allowing spoken instructions to potentially override the VOICE_SYSTEM_PROMPT.
[CREDENTIALS_UNSAFE]: The skill accesses sensitive API keys by reading configuration files from the local file system.
It specifically targets ~/.claude/skills/eleven-labs-skill/config.json to reuse an ElevenLabs API key, which is a form of cross-skill credential access.

voice-mode