View videos

A video contains both pictures and sound. Read the subtitles and look at the frames before drawing conclusions about the whole thing. Downloading a video, extracting images, or reading a list of filenames does not mean you have viewed it.

Start with the question

Work out what your human needs to know. A question about something the speaker said may only need the transcript. A question about a demonstration, design, movement, or something on screen needs visual inspection.

Accept a URL or an existing local file. Honour any timestamp or range the human gives you. For a long video, use the transcript and an overview to find the relevant sections, then inspect those sections more closely. Do not silently treat a sparse sample as complete coverage.

Use yt-dlp for remote videos and ffmpeg / ffprobe for local processing. Prefer these tools to opening a browser and manually scrubbing a player. This should run silently without opening windows or playing sound through the human's speakers.

Prepare the video

Use the bundled helper rather than rewriting the extraction and subtitle parsing each time. Resolve the absolute directory containing this SKILL.md, then run its scripts/view_video.py. Python 3.10+, FFmpeg, and ffprobe are required; URLs also need yt-dlp. Pillow is only needed for optional contact sheets.

Before downloading, check yt-dlp --version and ensure the executable the helper will use is up to date. Video sites change frequently, so an installed but outdated yt-dlp may no longer work. Update through its original installation method using the official update guidance, then confirm the version on PATH. Check this first when extraction or download errors occur, before assuming the video is unavailable or adding workarounds.