ec-watch

/ec-watch — Claude watches a video

You don't have a video input; this skill gives you one. A Python script downloads the video, extracts frames as JPEGs, gets a timestamped transcript (native captions first, then local faster-whisper fallback), and prints frame paths. You then analyze each frame path using the MiniMax understanding_image MCP tool to see the images and combine them with the transcript to answer the user.

Step 0 — First-run setup (runs once only)

The very first time /ec-watch is invoked, setup runs automatically and sets up faster-whisper, verifies binaries, and marks setup complete. After that, Step 0 is skipped entirely — subsequent invocations go straight to Step 1.

To re-run setup manually (only if needed):

python "C:/Users/Emir/.claude/skills/ec-watch/scripts/setup.py"

When to use

• User pastes a video URL (YouTube, Vimeo, X, TikTok, Twitch clip, most yt-dlp-supported sites) and asks about it.
• User points at a local video file (.mp4, .mov, .mkv, .webm, etc.) and asks about it.
• User types /ec-watch <url-or-path> [question].

Recommended limits

• Best accuracy: videos under 10 minutes. Frame coverage scales inversely with duration.
• Hard caps: 100 frames total and 2 fps. Token cost grows with frame count:
  • ≤30s → ~1-2 fps (up to 30 frames)
  • 30s-1min → ~40 frames
  • 1-3min → ~60 frames
  • 3-10min → ~80 frames
  • >10min → 100 frames, sparsely spaced (warning printed)
• For long videos, ask if they want to focus on a specific section.

How to invoke

Step 1 — parse the user input. Separate the video source (URL or path) from any question the user asked. Example: /ec-watch https://youtu.be/abc what language is this in? → source = https://youtu.be/abc, question = what language is this in?.

Step 2 — run the watch script. Use python on Windows:

python "C:/Users/Emir/.claude/skills/ec-watch/scripts/watch.py" "<source>"

Optional flags:

• --start T / --end T — focus on a section (SS, MM:SS, or HH:MM:SS)
• --max-frames N — lower the cap (default 80, max 100)
• --resolution W — frame width in px (default 512)
• --fps F — override auto-fps (clamped to 2 fps max)
• --out-dir DIR — working directory (default: tmp)
• --whisper local — force local faster-whisper (default, no API key needed)
• --no-whisper — disable Whisper fallback (frames-only)

Examples:

# Last 10 seconds
python "C:/Users/Emir/.claude/skills/ec-watch/scripts/watch.py" video.mp4 --start 50 --end 60

# Zoom into 2:15 → 2:45
python "C:/Users/Emir/.claude/skills/ec-watch/scripts/watch.py" "$URL" --start 2:15 --end 2:45

# From 1h12m to end
python "C:/Users/Emir/.claude/skills/ec-watch/scripts/watch.py" "$URL" --start 1:12:00

Step 3 — Analyze frames with MiniMax MCP.

After frames are extracted, use the MiniMax understanding_image MCP tool to analyze each frame. The tool name is mcp__minimax__understanding_image and it accepts:

• prompt: "Describe what's on screen, any UI elements, text visible, and the overall context."
• image_url: The file path to the JPEG frame

Process frames in batches of 5-10 for efficiency. Combine the visual analysis with the transcript to answer the user.

Step 4 — answer the user. You now have:

• Frames — what's on screen at each timestamp
• Transcript — what's said at each timestamp (via whisper (local) or captions)

If the user asked a specific question, answer citing timestamps. If they didn't ask, summarize the video — key moments, visuals, spoken content.

Step 5 — clean up. Delete the working directory if user won't ask follow-ups:

rm -rf "C:/Users/Emir/AppData/Local/Temp/watch-xxx"

Transcription

The script gets a timestamped transcript in one of two ways:

1. Native captions (free, preferred). yt-dlp pulls subtitles from the source.
2. Local faster-whisper fallback (free, no API key). Extracts audio and transcribes locally using faster-whisper on CPU.

No API key needed for local transcription.

Failure modes

• Setup preflight failed → run python "C:/Users/Emir/.claude/skills/ec-watch/scripts/setup.py"
• No transcript → captions missing AND Whisper unavailable. Proceed frames-only.
• Long video warning → offer to re-run focused on a section with --start/--end
• Download fails → tell the user plainly; do not retry

Token efficiency

• 80 frames at 512px ≈ 50-80k image tokens
• Transcript is cheap (a few thousand tokens max)
• Only bump resolution to 1024 when user needs to read on-screen text

If user asks a follow-up, do not re-run the script — you already have frames and transcript in context.

Security

• Video downloaded locally via yt-dlp
• Audio transcribed locally via faster-whisper (no data leaves your machine)
• No API keys stored or transmitted
• Working directory auto-deleted when done