Ask an AI Agent questions about a video without treating it as continuous viewing

Claude Video Watch Skill analyzes a public video URL or local file through the /watch command. It collects captions, uses Whisper only when needed, and samples still frames by scene. Claude or a compatible Agent receives text plus visual timestamps rather than a continuously played video.

Useful inputs and outputs

Give it a URL or local path, ideally with a time range when the question concerns one segment. It can summarize YouTube videos or local recordings, locate a point where an idea appears, compare speech with the image, and answer questions about tutorials, product demos, and recorded presentations. You can choose transcript, efficient, balanced, or higher-frame-detail modes to trade visual coverage against token and processing cost.

For a first check, use a public-captioned video of one to five minutes: verify yt-dlp access, request a low-detail summary with three timestamps, compare it against the video, then separately test an uncaptained video with Whisper fallback.

Setup and data boundary

Claude Code can install it through the upstream plugin marketplace. Codex and other Agent-Skills-compatible tools can use the repository's general installation approach for watch. ffmpeg and yt-dlp are required. macOS can install through Homebrew on first run; Windows and Linux show commands that still require user confirmation.

Native captions avoid an API key. If none are available, Groq or OpenAI Whisper can receive audio, with cost and privacy implications. Disabling Whisper permits frame-only analysis but substantially weakens understanding of speech.

When not to rely on it

Sampled frames can miss quick interfaces, short actions, background details, and audio-only events. Bad captions also corrupt summaries and timestamps. This is for research, summarization, and locating material, not frame-accurate safety forensics, medical interpretation, or legal evidence analysis.

Source and license