Claude Video: The 7.8k-Star Free Skill That Lets Claude Watch YouTube
Claude can read your code, browse the web, and write your docs β but until now, paste a YouTube link and it could only guess from the title. Claude Video changes that. It's a free, MIT-licensed skill (7,800+ stars on GitHub) that lets Claude actually watch a video: it extracts frames, pulls the transcript, and answers questions grounded in what's really on screen and in the audio.
Ask "What happens at the 30-second mark?" or "Why did this video go viral?" and Claude answers from the visuals and the audio β not from metadata.
How It Works
The skill does two things under the hood:
1. Frame extraction. Using ffmpeg, it pulls frames from the video and feeds them to Claude as images. Four detail modes let you trade speed against depth:
| Mode | What it does | Speed |
|---|---|---|
transcript | Text-only (captions) | ~4.5s |
efficient | Keyframes only, 50-frame cap | ~0.5s |
balanced | Scene-change detection, 100-frame cap | ~21s |
token-burner | Scene-change detection, uncapped | Full coverage |
A smart deduplication pass drops near-identical frames (held slides, static screens), so your frame budget is spent on distinct content β the threshold is even tuned for code diffs and scrolling terminals.
2. Transcription. It grabs native captions via yt-dlp first β free and instant. Only when captions don't exist (local files, some TikToks) does it fall back to the Whisper API, preferring Groq's whisper-large-v3 over OpenAI's for cost and speed. For most YouTube videos, you need zero API keys.
Usage
/watch https://youtu.be/dQw4w9WgXcQ what happens at the 30 second mark?
/watch ~/Movies/screen-recording.mp4 when does the UI break?
/watch https://vimeo.com/123 --start 2:15 --end 2:45
The --start/--end flags are the power move: focusing on a time range gives you denser frame coverage at lower token cost. Other useful flags: --resolution (bump to 1024px to read on-screen text), --max-frames, and --timestamps.
Installation
For Claude Code:
/plugin marketplace add bradautomates/claude-video
/plugin install watch@claude-video
It also works on Codex, Cursor, Copilot, Gemini CLI and 50+ other agents via npx skills add bradautomates/claude-video -g, and on claude.ai web through a downloadable .skill bundle. Dependencies (ffmpeg, yt-dlp) auto-install on first run.
What It's Good For
- β’Summarizing long videos β get the key points of a 40-minute talk without watching it
- β’Studying viral content β analyze pacing, hooks, and visuals of trending videos
- β’Debugging screen recordings β "when does the UI break?" on a local
.mp4 - β’Analyzing tutorials β extract the exact steps and commands without scrubbing
It supports YouTube, TikTok, Instagram, Loom, X, Vimeo and 100+ sources, plus local .mp4/.mov/.mkv/.webm files.
Limitations Worth Knowing
Capped modes spread frames across the full duration, so coverage gets sparse past ~10 minutes β use token-burner or a --start/--end range for long videos. And higher resolution or uncapped frames directly increase image token consumption, so pick the mode that matches your question.
The Takeaway
Video was the last big blind spot for AI coding assistants. A single open-source skill β built by Brad Bonanno of Solaris Automation β closes it with a pragmatic design: captions first, frames deduplicated, tokens spent only where they matter. If your workflow ever touches YouTube links or screen recordings, this is a one-command install that pays for itself the first time you skip a 30-minute video.
Reference: github.com/bradautomates/claude-video
#ClaudeCode #AITools #VideoAnalysis #OpenSource #ClaudeSkills #Automation
βοΈ The Author: Do Ngoc Hoan Founder of CookConnects.ca & Wizy.ca. Bridging the gap between advanced algorithms and business execution. I write for technical founders looking to scale their impact with AI and robust engineering.
Hoan Do
Founder at Wizy Marketing Agency. Passionate about helping Vietnamese businesses in North America scale with modern technology and premium marketing strategies.
Learn more about us β