letmewatch

Let Me Watch

Analyze video content by extracting key frames using ffmpeg scene detection and viewing them as images.

When to Activate

• User asks you to watch, review, or analyze a video
• User shares a video file path (.mp4, .mov, .mkv, .webm, .avi)
• User wants feedback on a screen recording or bug report video
• User runs /letmewatch:video, /letmewatch:video-last, or /letmewatch:video-dir

Prerequisites

• ffmpeg (required): brew install ffmpeg (macOS) or apt install ffmpeg (Linux)
• whisper (optional, for audio): pip install openai-whisper or pip install mlx-whisper

How It Works

1. The extraction script uses ffmpeg scene detection to find frames where visuals change significantly (like LogRocket focusing on user interactions)
2. Each frame is timestamped (frame_01m23s.jpg) so you can reference specific moments
3. Frames are resized to 720p JPEG to stay within context limits
4. Audio is transcribed if whisper is installed
5. You view frames in batches of 8, building a complete understanding of the video

Processing a Video

Step 1: Extract frames

Run the extraction script bundled with this skill:

python3 ${CLAUDE_PLUGIN_ROOT}/skills/letmewatch/video-extract.py "<video_path>"

Read the output to find:

• WORK_DIR — where frames are stored
• TOTAL_FRAMES — how many frames were extracted
• TRANSCRIPT — path to audio transcript (or "none")
• FRAMES — list of frame file paths

Step 2: Read transcript (if available)

If TRANSCRIPT is not "none", read the transcript file first for audio/narration context.

Step 3: View frames in batches

Read frames in batches of 8 using the Read tool (all 8 in parallel). For each batch:

• Note the timestamp in each filename (e.g., frame_00m23s.jpg = 0 minutes 23 seconds)
• Describe what you observe: UI state, user interactions, changes between frames, errors

Step 4: Synthesize

After viewing all frames, provide a timestamped summary. Tailor your response:

• Bug/UI review: Identify the issue, when it occurs, suggest fixes
• Screen recording: Describe the workflow and any issues spotted
• Tutorial/walkthrough: Summarize concepts and key takeaways
• General video: Describe content and answer questions

Step 5: Cleanup

Remove the temp directory:

rm -rf <WORK_DIR>

Important Notes

• Always reference timestamps: "At 01m15s, the error dialog appears"
• If scene detection yields few frames, the script auto-falls back to interval extraction
• The scene detection threshold is 0.1 by default (catches most UI changes)
• Max 40 frames per video, batched in groups of 8
• If frames look similar, note the video was static in that range