What problem does it solve? Video transcription alone loses critical information shown on screen—slides, terminal output, diagrams, and code. This Skill combines spoken transcripts with sequential visual frame analysis via a Vision-Language Model to produce a complete, searchable markdown knowledge base from any YouTube video. ## Core Features & Use Cases - Multimodal Extraction: Downloads auto-subtitles (.vtt) and extracts 8 sequential frames per chapter with ffmpeg, then fuses both through a VLM. - Semantic Chunking: Splits videos by chapters or 3-minute chunks to stay within context limits while preserving structure. - Focusable Analysis: Accepts an optional system prompt addendum to steer the VLM toward specific topics like CVEs, architectures, or methodologies. - Use Case: Feed a 2-hour security conference talk URL and receive a chapter-by-chapter markdown document capturing both the narration and every diagram shown on screen, ready to be fused into research notes. ## Quick Start Run the omni_video_extractor.py script with a YouTube URL, an output directory, and your .env credentials file to generate a markdown knowledge base of the video.
