The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses the subprocess.run function to invoke ffmpeg and ffprobe for audio validation and preprocessing. While the code follows best practices by using list-style arguments to prevent shell injection, it still performs direct system command execution.
[EXTERNAL_DOWNLOADS]: The skill downloads pre-trained speaker diarization models from Hugging Face (pyannote/speaker-diarization-3.1) at runtime using the pyannote.audio library.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted audio data, transcribes it, and then prepares a payload specifically designed for handoff to downstream LLM agents. This creates a risk where instructions spoken within the audio could influence the behavior of the subsequent agent.
Ingestion points: Audio files are ingested and processed through the validate_audio_file and transcribe_chunk functions in SKILL.md.
Boundary markers: The skill uses speaker labels and timestamps to structure the output, but it does not employ explicit delimiters or instructions to ignore commands embedded in the transcript text.
Capability inventory: The skill has the ability to perform network POST requests via httpx.post and execute shell commands via subprocess.run as defined in SKILL.md.
Sanitization: A normalize_transcript function is included to clean model artifacts (like capitalization or noise markers), but it does not perform sanitization for prompt injection patterns.
[DATA_EXFILTRATION]: The skill includes a function post_transcript_to_cms that sends transcribed content and associated metadata to external CMS endpoints via httpx. While the endpoint and credentials are provided as parameters, this represents a standard network exfiltration path.

Voice AI Integration Engineer