claude-video: Empowering Claude to Watch and Analyze Any Video Content

Introduction

Claude, a powerful AI, excels at reading webpages, running scripts, and browsing repositories. However, its native ability to "watch" and comprehend video content is limited. It often relies on titles or incomplete transcripts, missing crucial visual and auditory context.

The claude-video repository introduces /watch, a groundbreaking tool that empowers Claude to truly understand any video. By simply providing a URL or local path, /watch automates the process of downloading, extracting scene-aware frames, and generating a timestamped transcript. This rich, multimodal data is then handed to Claude, allowing it to answer questions and provide insights grounded in what it has actually seen and heard in the video.

For example, you can ask:

/watch https://youtu.be/dQw4w9WgXcQ what happens at the 30 second mark?

Installation

Installing claude-video is straightforward across various platforms and AI agents.

For Claude Code:

/plugin marketplace add bradautomates/claude-video
/plugin install watch@claude-video

For Codex, Cursor, Copilot, Gemini CLI, and other Agent Skills hosts:

npx skills add bradautomates/claude-video -g

The -g flag installs globally for your user, making it available across all projects.

For claude.ai (web interface):

Download the watch.skill file from the latest release.
Go to Claude's Settings, then Capabilities, then Skills.
Click the + button and upload the downloaded watch.skill file.

Ensure "Code execution and file creation" is enabled under Capabilities, as the skill uses external tools like ffmpeg and yt-dlp.

First Run Setup:

On its first execution, /watch will automatically check for ffmpeg and yt-dlp and guide you through their installation if needed, for example, via brew on macOS, or printing commands for Linux/Windows. It also helps set up a Whisper API key for transcription fallback, though native captions cover most public videos for free.

Examples

claude-video opens up a wide range of practical applications for video analysis:

Analyze Content: Understand the structure or hooks of a viral video.

/watch https://youtu.be/<viral-video> what hook did they open with?

Diagnose Bugs: Get Claude to watch a screen recording and identify issues.
```
/watch bug-repro.mov what's going wrong?
```
Summarize Videos: Quickly get the key moments and structure of a long video.
```
/watch https://youtu.be/<long-thing> summarize this
```

Filter Hype: Extract the core information from launch videos.

/watch https://youtu.be/<launch-video> what's actually new, skip the hype

Create Notes from Playlists: Turn a series of videos into searchable summaries.
```
/watch https://youtu.be/<video> summarize this to a note
```

You can also focus on specific sections for denser analysis and lower token cost:

/watch https://youtu.be/abc --start 2:15 --end 2:45

Why Use It

claude-video transforms Claude into a truly multimodal assistant, capable of understanding video content in a way that was previously impossible.

Deep Understanding: Claude doesn't just guess from titles, it "sees" the frames and "hears" the audio, providing answers grounded in the actual content.
Efficiency: Get summaries, analyses, and answers from videos much faster than watching them manually, even at accelerated speeds.
Versatility: Supports a vast array of video sources, including YouTube, Loom, TikTok, X, Instagram, and local files, thanks to yt-dlp.
Cost-Effective: Leverages free native captions whenever available, with Whisper API as an intelligent fallback for videos without captions. The tool also manages frame budgets to optimize token usage.
Practical Applications: From content analysis and bug diagnosis to efficient summarization and note-taking, /watch offers tangible benefits for developers, researchers, and content creators.