video-understand

Understand video content locally using ffmpeg for frame extraction and Whisper for transcription. Fully offline, no API keys required.

Prerequisites

• ffmpeg + ffprobe (required): brew install ffmpeg
• openai-whisper (optional, for transcription): pip install openai-whisper

Commands

bash
# Scene detection + transcribe (default)
python3 skills/video-understand/scripts/understand_video.py video.mp4

# Keyframe extraction
python3 skills/video-understand/scripts/understand_video.py video.mp4 -m keyframe

# Regular interval extraction
python3 skills/video-understand/scripts/understand_video.py video.mp4 -m interval

# Limit frames extracted
python3 skills/video-understand/scripts/understand_video.py video.mp4 --max-frames 10

# Use a larger Whisper model
python3 skills/video-understand/scripts/understand_video.py video.mp4 --whisper-model small

# Frames only, skip transcription
python3 skills/video-understand/scripts/understand_video.py video.mp4 --no-transcribe

# Quiet mode (JSON only, no progress)
python3 skills/video-understand/scripts/understand_video.py video.mp4 -q

# Output to file
python3 skills/video-understand/scripts/understand_video.py video.mp4 -o result.json

CLI Options

| Flag | Description | |------|-------------| | video | Input video file (positional, required) | | -m, --mode | Extraction mode: scene (default), keyframe, interval | | --max-frames | Maximum frames to keep (default: 20) | | --whisper-model | Whisper model size: tiny, base, small, medium, large (default: base) | | --no-transcribe | Skip audio transcription, extract frames only | | -o, --output | Write result JSON to file instead of stdout | | -q, --quiet | Suppress progress messages, output only JSON |

Extraction Modes

| Mode | How it works | Best for | |------|-------------|----------| | scene | Detects scene changes via ffmpeg select='gt(scene,0.3)' | Most videos, varied content | | keyframe | Extracts I-frames (codec keyframes) | Encoded video with natural keyframe placement | | interval | Evenly spaced frames based on duration and max-frames | Fixed sampling, predictable output |

If scene mode detects no scene changes, it automatically falls back to interval mode.

Output

The script outputs JSON to stdout (or file with -o). See references/output-format.md for the full schema.

json
{
  "video": "video.mp4",
  "duration": 18.076,
  "resolution": {"width": 1224, "height": 1080},
  "mode": "scene",
  "frames": [
    {"path": "/abs/path/frame_0001.jpg", "timestamp": 0.0, "timestamp_formatted": "00:00"}
  ],
  "frame_count": 12,
  "transcript": [
    {"start": 0.0, "end": 2.5, "text": "Hello and welcome..."}
  ],
  "text": "Full transcript...",
  "note": "Use the Read tool to view frame images for visual understanding."
}

Use the Read tool on frame image paths to visually inspect extracted frames.

References

• references/output-format.md -- Full JSON output schema documentation