The gist

Claude Video is a script and AI assistant skill created by Brad Bonanno. It solves the problem of language models like Claude being unable to process video directly. The tool allows users to provide a video URL or local file, enabling the AI to analyze both the visual frames and spoken audio transcript to answer questions, effectively letting it 'watch' the video.

What it does

Analyze video from URLs (YouTube, TikTok, Loom, etc.) or local files (.mp4, .mov).
Extract timestamped transcripts using free native captions when available.
Transcribe audio using Whisper via Groq or OpenAI APIs as a fallback for caption-less video.
Select and extract relevant frames based on scene changes or efficient keyframes.
Summarize meetings, diagnose bugs from screen recordings, or deconstruct ad creative.
Focus analysis on specific time windows (e.g., --start 2:15 --end 2:45) for greater detail.

How it works

A user runs the /watch command with a video link or file path and a question. The script uses yt-dlp and ffmpeg to check for captions, download content if needed, and extract frames. It passes the visual frames and the timestamped transcript to Claude's multimodal vision model for analysis. The tool is an open-source CLI skill that integrates with multiple agent hosts and requires a one-time setup for its dependencies.

Best for

This tool is ideal for a developer diagnosing a bug from a screen recording or a content strategist analyzing the opening hook of a competitor's video without having to manually watch and transcribe it.

Watch out for

When analyzing videos longer than 10 minutes, the default modes provide a 'sparse scan' that may miss details. For full coverage on longer clips, you must either re-run the command focused on a specific time segment or use the more token-intensive token-burner detail mode.