video-watching

To generate a time-stamped transcript with visual descriptions, provide a video path or URL and the skill processes it using FFmpeg and Whisper-based tooling. It outputs an aligned transcript alongside a frames index with inline on-screen visual descriptions.

You can use yt-dlp to supply online video URLs for transcript extraction. The skill processes the URL to produce a time-stamped transcript and a frames index with inline visual descriptions stored locally.

Running Whisper-based video transcription locally requires ffmpeg, ffprobe, yt-dlp, whisper-cli, whisper-cpp, and python3. These dependencies extract frames and align speech to produce the time-stamped transcript.

The best way to analyze marketing or product demo videos without watching every frame is generating a time-stamped transcript with inline on-screen visual descriptions. This enables replayable reviews and knowledge extraction for AI and human reasoning.

The video pipeline supports extracting optional anchor frames for opening and closing hooks. It aligns these anchor frames with the time-stamped transcript to provide visual context for review.

Transcript and frame timeline misalignment occurs when ffprobe or ffmpeg fails to extract accurate timestamps. The skill relies on these dependencies to properly synchronize the frames.json index with the Whisper-generated speech timeline.