Watch Video

What This Skill Does

1. Downloads the media (if URL) or reads the local file
2. Extracts key frames and audio from videos
3. Uses Gemini Flash to analyze visual content (UI state, text, user actions)
4. Transcribes audio narration (if present)
5. Synthesizes everything into structured notes with:
   • What's shown and what's happening
   • Exact text visible on screen (errors, URLs, labels, code)
   • What was said (audio transcript)
   • Key observations and actionable takeaways
   • Confidence level and clarifying questions

Setup

pip install eyeroll
eyeroll init          # or: export GEMINI_API_KEY=your-key

pip install eyeroll
ollama serve          # start Ollama

brew install yt-dlp

Commands

Analyze a video or screenshot

eyeroll watch <source> [--context "..."] [--backend gemini|ollama] [--model MODEL] [--max-frames 20] [--output report.md] [--verbose]

• source

  : URL (Loom, YouTube, any yt-dlp supported site) or local file path (.mp4, .webm, .mov, .png, .jpg, .gif)

• --context

  : Text context from the person who shared the video — Slack message, issue description, PR reference. Significantly improves analysis quality.

• --max-frames

  : Maximum frames to extract from video (default: 20)

• --output

  : Write notes to file instead of stdout

• --verbose

  : Show progress details

When To Use This Skill

• User shares a video or screenshot and wants the agent to understand it
• User pastes a Loom or YouTube link
• User says "watch this", "look at this recording", "look at this video"
• User asks to fix a bug based on a video or screenshot
• User asks to build something based on a demo video
• User asks to create a skill/plugin/subagent from a tutorial video
• User shares a screen recording with context about what to do with it
• User asks to generate notes, a report, or documentation from a video

Example Interactions

eyeroll watch https://loom.com/share/abc123

eyeroll watch <video_path> --context "checkout broken after billing migration merge"

eyeroll watch <url>

eyeroll watch <url> --context "create a skill based on this tutorial"

eyeroll watch screenshot.png

eyeroll watch <url>

Rules

• Always check that GEMINI_API_KEY is set before running analysis.
• If ffmpeg is not on PATH, the bundled imageio-ffmpeg fallback is used.
• For videos under 20MB and 2 minutes, the full video is sent to Gemini directly (more accurate). Longer videos use frame-by-frame analysis.
• When context is available (--context), always include it — it dramatically improves quality, especially for silent recordings.
• Do NOT hallucinate text, error messages, or code. Only report what is actually visible/audible.
• If the video is ambiguous or unclear, output clarifying questions rather than guessing.
• The notes are a starting point — when the user asks to take action (fix, build, create), use the notes plus codebase context to do so.
• Warn the user if yt-dlp is not installed when they provide a URL.

Supported Input Types

Cost

• Short video (<2min, direct upload): ~$0.01-0.05 per analysis
• Long video (frame-by-frame, 20 frames): ~$0.02-0.10 per analysis
• Audio transcription: ~$0.01 per minute
• Synthesis: ~$0.01 per report