The Agent Skills Directory

[DYNAMIC_EXECUTION]: The setup.sh script generates a sitecustomize.py file within the Python virtual environment's site-packages. This script dynamically patches torch.load to set weights_only=False. This effectively disables a security feature in PyTorch designed to prevent arbitrary code execution when loading untrusted model files, although it is intended here to ensure compatibility with specific diarization models.- [EXTERNAL_DOWNLOADS]: The skill downloads AI models from HuggingFace (a well-known repository) during the transcription and diarization phases. It also uses yt-dlp to download video content from YouTube when a URL is provided by the user.- [COMMAND_EXECUTION]: The skill makes extensive use of system shell commands and subprocesses to run media utilities including ffmpeg, ffprobe, yt-dlp, pandoc, and weasyprint. It also executes Python scripts for transcription and frame analysis.- [DATA_EXFILTRATION]: The ingest.sh script is configured to access local browser cookies (defaulting to Google Chrome) via yt-dlp. This is used to authenticate with YouTube for downloading restricted or bot-protected content, but it involves reading sensitive local data from the user's browser profile.- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data from YouTube videos and local files. It transcribes audio and extracts visual frames which are then processed by the AI agent. Ingestion points: Video URLs or local files processed via ingest.sh and transcribe.sh. Boundary markers: None identified for transcript or frame context injection into the main prompt. Capability inventory: Subprocess execution (ffmpeg, yt-dlp), file system writes, and PDF report generation. Sanitization: No explicit sanitization of transcribed text or visual metadata before it is presented to the agent for analysis.

video-analyzer