claude-vision

claude-vision

Visual context skills for Claude Code. Gives Claude the ability to see your clipboard, screenshots, and videos.

Why This Exists

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

Video Analysis

Claude doesn't natively support video. But with some clever extraction, we can give it Gemini-like video understanding:

1. YouTube videos → Extract auto-captions + key frames
2. Local videos → Scene detection for key frames + optional whisper transcription
3. 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.

What's Included

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

Quick Start

# 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

Docker Images

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.

GPU Acceleration (Full image only)

Whisper transcription speed:

Setup	Speed	1 hour video
CPU	~10x realtime	~6 minutes
GPU (CUDA)	~50x realtime	~1 minute

To enable GPU:

1. NVIDIA GPU + drivers installed
2. Install nvidia-container-toolkit
3. cv-run will auto-detect and enable GPU

Platform Support

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

Windows Users: Why WSL?

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.

Prerequisites

Clipboard/screenshot:

• WSL: PowerShell (built-in)
• macOS: brew install pngpaste (for images)
• Linux: apt install xclip

Video processing: Docker (required)

Usage

Clipboard

/clipboard what is this?
/clipboard explain this code
/clipboard describe what you see

Screenshot

Auto-detects your screenshot directory (configured during setup) and finds the most recent file by timestamp.