claude-configs/skills/youtube-transcriber/SKILL.md

name	description
youtube-transcriber	Transcribe YouTube videos of any length using OpenAI's GPT-4o-transcribe model. Supports parallel processing for multiple videos and automatic chunking for long content. USE WHEN user says 'transcribe video', 'transcribe youtube', 'get transcript', or provides YouTube URLs.

YouTube Transcriber - High-Quality Video Transcription

When to Activate This Skill

• "Transcribe this YouTube video"
• "Get a transcript of [URL]"
• "Transcribe these videos" (multiple URLs)
• User provides YouTube URL(s) needing transcription
• "Extract text from video"
• Any request involving YouTube video transcription

Script Location

Primary script: /mnt/NV2/Development/youtube-transcriber/transcribe.py

Key Features

• Parallel processing: Multiple videos can be transcribed simultaneously
• Unlimited length: Auto-chunks videos >10 minutes to prevent API limits
• Organized output:
  • Transcripts → output/ directory
  • Temp files → temp/ directory (auto-cleaned)
• High quality: Uses GPT-4o-transcribe by default (reduced hallucinations)
• Cost options: Can use -m gpt-4o-mini-transcribe for 50% cost savings

Basic Usage

Single Video

cd /mnt/NV2/Development/youtube-transcriber
uv run python transcribe.py "https://youtube.com/watch?v=VIDEO_ID"

Output: output/Video_Title_2025-11-10.txt

Multiple Videos in Parallel

cd /mnt/NV2/Development/youtube-transcriber

# Launch all in background simultaneously
uv run python transcribe.py "URL1" &
uv run python transcribe.py "URL2" &
uv run python transcribe.py "URL3" &
wait

Why parallel works: Each transcription uses unique temp files (UUID-based) in temp/ directory.

Cost-Saving Mode

cd /mnt/NV2/Development/youtube-transcriber
uv run python transcribe.py "URL" -m gpt-4o-mini-transcribe

When to use mini: Testing, casual content, bulk processing. Quality is the same as gpt-4o-transcribe but ~50% cheaper.

Command Options

uv run python transcribe.py [URL] [OPTIONS]

Options:
  -o, --output PATH           Custom output filename (default: auto-generated in output/)
  -m, --model MODEL          Transcription model (default: gpt-4o-transcribe)
                             Options: gpt-4o-transcribe, gpt-4o-mini-transcribe, whisper-1
  -p, --prompt TEXT          Context prompt for better accuracy
  --chunk-duration MINUTES   Chunk size for long videos (default: 10 minutes)
  --keep-audio              Keep temp audio files (default: auto-delete)

Workflow for User Requests

Single Video Request

1. Change to transcriber directory
2. Run script with URL
3. Report output file location in output/ directory

Multiple Video Request

1. Change to transcriber directory
2. Launch all transcriptions in parallel using background processes
3. Wait for all to complete
4. Report all output files in output/ directory

Testing/Cost-Conscious Request

1. Always use -m gpt-4o-mini-transcribe for testing
2. Mention cost savings to user
3. Quality is identical to full model

Example Responses

User: "Transcribe this video: https://youtube.com/watch?v=abc123"

Assistant Action:

cd /mnt/NV2/Development/youtube-transcriber
uv run python transcribe.py "https://youtube.com/watch?v=abc123"

Report: "✅ Transcript saved to output/Video_Title_2025-11-10.txt"

---

User: "Transcribe these 5 videos: [URL1] [URL2] [URL3] [URL4] [URL5]"

Assistant Action: Launch all 5 in parallel:

cd /mnt/NV2/Development/youtube-transcriber
uv run python transcribe.py "URL1" &
uv run python transcribe.py "URL2" &
uv run python transcribe.py "URL3" &
uv run python transcribe.py "URL4" &
uv run python transcribe.py "URL5" &
wait

Report: "✅ All 5 videos transcribed successfully in parallel. Output files in output/ directory"

Technical Details

How it works:

1. Downloads audio from YouTube (via yt-dlp)
2. Saves to unique temp file: temp/download_{UUID}.mp3
3. Splits long videos (>10 min) into chunks automatically
4. Transcribes with OpenAI API (GPT-4o-transcribe)
5. Saves transcript: output/Video_Title_YYYY-MM-DD.txt
6. Cleans up temp files automatically

Parallel safety:

• Each process uses UUID-based temp files
• No file conflicts between parallel processes
• Temp files auto-cleaned after completion

Auto-chunking:

• Videos >10 minutes: Split into 10-minute chunks
• Context preserved between chunks
• Prevents API response truncation

Requirements

• OpenAI API key: $OPENAI_API_KEY environment variable
• Python 3.10+ with uv package manager
• FFmpeg (for audio processing)
• yt-dlp (for YouTube downloads)

Check requirements:

echo $OPENAI_API_KEY  # Should show API key
which ffmpeg          # Should show path

Output Format

Transcripts are saved as plain text with metadata:

================================================================================
YouTube Video Transcript (Long Video)
================================================================================

Title: Video Title Here
Uploader: Channel Name
Duration: 45m 32s
URL: https://youtube.com/watch?v=VIDEO_ID

================================================================================

[Full transcript text with proper punctuation...]

Best Practices

1. Always use parallel for multiple videos - It's 6x faster
2. Use mini model for testing - Same quality, half the cost
3. Check output/ directory - All transcripts organized there
4. Temp files auto-clean - No manual cleanup needed
5. Add context prompts for technical content:

   uv run python transcribe.py "URL" \
     -p "Technical discussion about Docker, Kubernetes, microservices"
   

Troubleshooting

API Key Missing:

export OPENAI_API_KEY="sk-proj-your-key-here"

FFmpeg Not Found:

sudo dnf install ffmpeg  # Fedora/Nobara

Parallel Conflicts (shouldn't happen with UUID temps):

• Each process creates unique temp file in temp/
• If issues occur, check temp/ directory permissions

Cost Estimates (as of March 2025)

• 5-minute video: $0.10 - $0.20
• 25-minute video: $0.50 - $1.00
• 60-minute video: $1.20 - $2.40

Using mini model: Reduce costs by ~50%

Quick Reference

# Single video (default quality)
uv run python transcribe.py "URL"

# Single video (cost-saving)
uv run python transcribe.py "URL" -m gpt-4o-mini-transcribe

# Multiple videos in parallel
for url in URL1 URL2 URL3; do
  uv run python transcribe.py "$url" &
done
wait

# With custom output
uv run python transcribe.py "URL" -o custom_name.txt

# With context prompt
uv run python transcribe.py "URL" -p "Context about video content"

Directory Structure

/mnt/NV2/Development/youtube-transcriber/
├── transcribe.py       # Main script
├── temp/              # Temporary audio files (auto-cleaned)
├── output/            # All transcripts saved here
├── README.md          # Full documentation
└── pyproject.toml     # Dependencies

Integration with Other Skills

With fabric skill: Process transcripts after generation

# 1. Transcribe
uv run python transcribe.py "URL"

# 2. Process with fabric
cat output/Video_Title_2025-11-10.txt | fabric -p extract_wisdom

With research skill: Transcribe source videos for research

# Transcribe multiple research videos in parallel
# Then analyze transcripts for insights

Notes

• Script requires being in its directory to work correctly
• Always change to /mnt/NV2/Development/youtube-transcriber first
• Parallel execution is safe and recommended for multiple videos
• Use mini model for testing to save costs
• Output files automatically named with video title + date
• Temp files automatically cleaned after transcription