The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill creates an indirect prompt injection surface by processing external, untrusted audio data and transforming it into text for downstream LLM consumers (e.g., summarization agents).
Ingestion points: Audio files in various formats (wav, mp3, mp4, etc.) processed in SKILL.md through functions like validate_audio_file and transcribe_chunk.
Boundary markers: The instructions do not define boundary markers or explicit prompts to ignore embedded instructions within the generated transcript segments.
Capability inventory: The skill utilizes shell command execution via subprocess (ffmpeg, ffprobe) and network operations via httpx to deliver results to external CMS endpoints.
Sanitization: The provided normalize_transcript function focuses on formatting and noise reduction rather than filtering or escaping potential prompt injection content.
[COMMAND_EXECUTION]: The skill uses the subprocess module to execute ffmpeg and ffprobe for audio validation and preprocessing. This is standard functionality for an audio engineering skill and targets user-provided file paths locally.
[EXTERNAL_DOWNLOADS]: The skill references external machine learning models and libraries, including faster-whisper and Hugging Face's pyannote/speaker-diarization-3.1. These are well-known resources in the AI community and their use is consistent with the skill's stated purpose.

agency-voice-ai-integration-engineer