Visual context skills for Claude Code. Gives Claude the ability to see your clipboard, screenshots, and videos.
Claude Code can technically read images—you just paste in the file path and it'll analyze them. But the workflow is clunky:
- You screenshot an error, but now you need to find where your OS saved it
- You copy something to your clipboard, but there's no way to share it
- You want Claude to look at a video, but... you can't
This plugin fixes all of that. Instead of hunting for file paths, just:
/clipboard what's this?
/screenshot explain this error
/video summarize this tutorial
Claude doesn't natively support video. But with some clever extraction, we can give it Gemini-like video understanding:
- YouTube videos → Extract auto-captions + key frames
- Local videos → Scene detection for key frames + optional whisper transcription
- Screen recordings → Same scene detection, finds the moments that matter
The /video skill spawns a dedicated analysis agent with a fresh 200k context. This means:
- 100+ frames instead of a handful
- Full transcripts without truncation
- Thorough analysis without bloating your conversation
The result: Claude can "watch" videos by analyzing representative frames and reading the transcript. It handles tutorials, error recordings, meetings, and lectures surprisingly well.
| Skill | Description |
|---|---|
/clipboard |
Read text or images from your clipboard |
/screenshot |
Analyze your latest screenshot |
/video |
Analyze YouTube videos or local recordings |
/claude-vision-setup |
Interactive setup wizard |
# Add the marketplace and install the plugin
/plugin marketplace add ellyseum/claude-plugins
/plugin install claude-vision
# Run setup (auto-runs on first use of any skill)
/claude-vision-setup| Image | Size | Includes | Use Case |
|---|---|---|---|
| Lite | ~500 MB | ffmpeg, yt-dlp | YouTube, screen recordings |
| Full | ~10 GB | ffmpeg, yt-dlp, whisper | Local videos needing transcription |
Which should I choose?
- Lite - For most users. YouTube has auto-captions, screen recordings usually don't need audio.
- Full - Only if you have local videos where you need to transcribe speech.
Whisper transcription speed:
| Setup | Speed | 1 hour video |
|---|---|---|
| CPU | ~10x realtime | ~6 minutes |
| GPU (CUDA) | ~50x realtime | ~1 minute |
To enable GPU:
- NVIDIA GPU + drivers installed
- Install nvidia-container-toolkit
cv-runwill auto-detect and enable GPU
| Feature | WSL | macOS | Linux |
|---|---|---|---|
| Clipboard (text) | Yes | Yes | Yes (xclip) |
| Clipboard (image) | Yes | Yes (pngpaste) | Yes (xclip) |
| Screenshot | Yes | Yes | Yes |
| Video analysis | Yes | Yes | Yes |
This plugin runs through WSL (Windows Subsystem for Linux), not native Windows. Here's why:
Docker Desktop for Windows requires WSL2 anyway. If you have Docker installed on Windows, you already have WSL2. The plugin just runs inside that same Linux environment where Docker lives.
This keeps the codebase simple—one set of bash scripts instead of maintaining parallel PowerShell versions.
Need native Windows support? If you have a use case that requires native Windows commands or a non-Docker workflow, open an issue and we'll consider it.
Clipboard/screenshot:
- WSL: PowerShell (built-in)
- macOS:
brew install pngpaste(for images) - Linux:
apt install xclip
Video processing: Docker (required)
/clipboard what is this?
/clipboard explain this code
/clipboard describe what you seeAuto-detects your screenshot directory (configured during setup) and finds the most recent file by timestamp.
/screenshot what's the error here?
/screenshot # Describes what it sees
/screenshot analyze the last 3 screenshots
/screenshot compare these two screenshots# YouTube (uses auto-captions)
/video https://youtube.com/watch?v=xyz summarize this
# Local recordings
/video what went wrong in my screen recording
# Follow-up questions (cached)
/video follow-up what was the error message?
# Cache management
/video --list
/video --clearThe cv-run script manages Docker:
cv-run ffmpeg -i video.mp4 ... # Run ffmpeg in container
cv-run yt-dlp https://... # Run yt-dlp
cv-run whisper audio.mp3 # Run whisper (full image only)
# Get images (pick one)
cv-run --pull-lite # Pull lite from ghcr.io (fast)
cv-run --pull-full # Pull full from ghcr.io (fast)
cv-run --build-lite # Build lite locally (~1 min)
cv-run --build-full # Build full locally (~6 min)
# Container management
cv-run --status # Show detailed status
cv-run --stop # Stop container
cv-run --rm # Remove containerImages are hosted at ghcr.io/ellyseum/claude-vision.
Config file: ~/.claude/claude-vision/config.json
{
"mode": "docker",
"os": "wsl",
"image_variant": "lite",
"screenshot_dir": "/mnt/c/Users/yourname/Pictures/Screenshots",
"created": "2026-02-02T...",
"version": "1.0"
}claude-vision/
├── .claude-plugin/ # Plugin metadata
├── Dockerfile # Full image (ffmpeg, yt-dlp, whisper)
├── Dockerfile.lite # Lite image (ffmpeg, yt-dlp only)
├── README.md
├── bin/
│ └── cv-run # Docker/local command router
├── hooks/
│ └── session-start.sh
└── skills/
├── clipboard/
├── screenshot/
├── video/
└── claude-vision-setup/
# Start Docker daemon
sudo systemctl start docker
# Or open Docker Desktopcv-run --pull-lite # or --pull-full (recommended, fast)
cv-run --build-lite # or --build-full (if pull fails)# Check nvidia-smi works on host
nvidia-smi
# Install nvidia-container-toolkit
# See: https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/install-guide.html
# Recreate container
cv-run --rm
cv-run ffmpeg -version # Will recreate with GPUMake sure you're in a proper terminal, not a headless SSH session.
sudo apt install xclipbrew install pngpasteMIT