# Claude Video — Give Claude the Ability to Watch Any Video

> A Claude Code skill that enables Claude to watch and understand any video by downloading, extracting frames, transcribing audio, and feeding structured data to the agent.

## Install

Save in your project root:

# Claude Video — Give Claude the Ability to Watch Any Video

## Quick Use
```bash
# Install the Claude Code skill
git clone https://github.com/bradautomates/claude-video.git
cp -r claude-video/.claude/skills/video ~/.claude/skills/video
# Then in Claude Code use the /watch command
/watch https://www.youtube.com/watch?v=example
```

## Introduction
Claude Video is a Claude Code skill that gives Claude the ability to watch and understand video content. Using the `/watch` command, it downloads a video, extracts representative frames, transcribes the audio track, and feeds the structured visual and textual data into Claude's context. This bridges the gap between Claude's strong multimodal reasoning and the video format it cannot natively consume.

## What Claude Video Does
- Downloads video from YouTube and other supported sources via yt-dlp
- Extracts key frames at configurable intervals for visual understanding
- Transcribes audio using Whisper for full speech-to-text coverage
- Combines frames and transcript into a structured payload for Claude's context
- Supports the `/watch` slash command for seamless integration into coding sessions

## Architecture Overview
The skill is a Python pipeline triggered by the `/watch` command. yt-dlp downloads the video, ffmpeg extracts frames at specified intervals, and Whisper produces a timestamped transcript from the audio. The skill then assembles frames and transcript into a structured context block injected into Claude's conversation. All processing runs locally.

## Self-Hosting & Configuration
- Clone the repository and copy the skill directory into your Claude Code skills folder
- Requires Python with yt-dlp, ffmpeg, and openai-whisper installed locally
- Configure frame extraction interval and maximum frame count in the skill settings
- Whisper model size is configurable to trade off speed versus transcription accuracy
- All processing runs locally with no external API calls beyond the video download

## Key Features
- End-to-end video understanding from a single `/watch` command
- Combines visual frame analysis with full audio transcription
- Works with YouTube, Vimeo, and any source supported by yt-dlp
- Configurable frame sampling to balance detail against context window usage
- Fully local processing with no data sent to third-party services

## Comparison with Similar Tools
- **Manual screenshots** — requires human effort; Claude Video automates frame extraction end to end
- **YouTube transcript copy-paste** — text only; Claude Video adds visual frame understanding
- **Video summarizer APIs** — cloud-dependent; Claude Video processes everything locally
- **Multimodal model native video** — not yet available in Claude; this skill bridges that gap

## FAQ
**Q: What video sources are supported?**
A: Any source supported by yt-dlp, including YouTube, Vimeo, and direct video URLs.

**Q: Does it send video data to external services?**
A: No. All frame extraction and transcription happen locally on your machine.

**Q: How many frames does it extract?**
A: This is configurable. The default extracts one frame every few seconds, capped at a maximum to fit Claude's context window.

**Q: Does it require a GPU for transcription?**
A: Whisper can run on CPU, though a GPU significantly speeds up transcription for longer videos.

## Sources
- https://github.com/bradautomates/claude-video

---
Source: https://tokrepo.com/en/workflows/asset-f189c294
Author: AI Open Source