The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/voice_persona.py invokes external system utilities using subprocess.run. It calls ffmpeg and afconvert for audio format transcoding (e.g., converting to PCM or WAV) and the say command on macOS for demonstration purposes. These invocations use list-based arguments which mitigate shell injection risks.
[INDIRECT_PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection when the optional LLM enhancement feature is used in scripts/voice_persona.py.
Ingestion points: User-supplied text is passed to the personalize function and subsequently the _llm_personalize function.
Boundary markers: The text is directly interpolated into a natural language prompt template without the use of clear delimiters or instructions for the model to ignore embedded commands.
Capability inventory: The skill environment has access to subprocess execution (ffmpeg, afconvert, say), local file system writes (audio and temporary pcm files), and network access (Edge TTS and LLM APIs).
Sanitization: The input text is not sanitized or escaped before being placed into the prompt string.
[EXTERNAL_DOWNLOADS]: The skill performs legitimate network operations to provide its core services. It communicates with Microsoft's Edge TTS service to synthesize voice responses and optionally with the DeepSeek API for persona-based text transformation. Additionally, the faster-whisper library downloads pre-trained transcription models from remote repositories during initial execution.

voice-persona