Video Chunking Skill

Process a D&D session video file through the complete chunking pipeline.

What This Skill Does

This skill orchestrates the complete video processing workflow:

1. Extract Audio: Uses FFmpeg to extract audio from the video file
2. Transcribe: Uses Whisper to transcribe the audio to text with speaker diarization
3. Classify: Uses Ollama to classify dialogue as in-character (IC) or out-of-character (OOC)
4. Extract Knowledge: Identifies NPCs, locations, quests, items, and other campaign elements
5. Chunk: Splits content into semantic segments based on topics and speaker changes
6. Generate Outputs: Creates data files, transcripts, and metadata

Prerequisites

Before running this skill, ensure:

• FFmpeg is available and accessible
• Ollama is running locally with an appropriate model
• PyAnnote speaker diarization models are downloaded
• Input video file path is provided
• Sufficient disk space for processing

Usage

When invoked, this skill will:

1. Verify all dependencies are available using the check_pipeline_health MCP tool
2. Run the processing pipeline via python cli.py process <video_file>
3. Monitor progress and report status
4. Handle errors gracefully and provide diagnostic information

Example Invocations

• "Process the session video at videos/session_001.mp4"
• "Chunk this D&D recording: recordings/adventure_ep5.mkv"
• "Transcribe and analyze this session file"

Command Reference

# Process a video file
python cli.py process <video_file> [--party default] [--session-id custom_id]

# Using Gradio UI
python app.py

Error Handling

Common issues and solutions:

• FFmpeg not found: Check FFmpeg installation with ffmpeg -version
• Ollama connection failed: Ensure Ollama is running (ollama serve)
• Out of memory: Process shorter videos or reduce batch sizes
• Transcription errors: Check audio quality and Whisper model availability
• Diarization failures: Verify PyAnnote models are downloaded

Output

The skill produces:

• Session data JSON with segments, speakers, and classifications
• Campaign knowledge JSON with NPCs, locations, quests, etc.
• Transcript files in human-readable format
• Processing logs with timestamps and diagnostics
• Output directory: output/YYYYMMDD_HHMMSS_sessionid/