YouTube Video Generation Pipeline

Channel: Głęboki Odczyt | Request episode: Create issue with whitepaper link

Generate YouTube videos from NotebookLM podcasts about computer science milestone papers - covering LLMs, distributed systems, networking, operating systems, and security.

Directory Structure

youtube-whitepapers/
├── notebooklm-automator/       # Browser automation for NotebookLM
│   ├── src/                    # Python source code
│   ├── mise.toml               # Task runner configuration
│   └── README.md               # Full automation docs
├── scripts/                    # Python automation scripts
│   ├── compress_images.py      # Batch PNG compression (>threshold)
│   ├── generate_status.py      # Generate status report for tracking
│   ├── generate_video.py       # Generate video from concat.txt + audio
│   ├── prepare_slides.py       # Extract/normalize slides from PDF
│   ├── rename_thumbnails.py    # Rename thumbnails to match whitepapers
│   ├── transcribe.py           # Batch transcription with Whisper
│   └── verify_video.py         # Verify video quality
├── tests/                      # Pytest test suite
├── mise.toml                   # Task runner configuration
├── whitepapers/
│   ├── distributed-computing/  # Distributed systems papers
│   ├── llm/                    # LLM research papers
│   ├── networking/             # Networking protocols & systems
│   ├── operating-systems/      # OS research papers
│   └── security/               # Security research papers
└── youtube/
    ├── pl/                     # Polish language assets
    │   ├── audio/              # NotebookLM podcast audio (.m4a)
    │   ├── slides/             # Presentation PDFs + extracted PNGs
    │   └── transcripts/        # Whisper transcriptions (.json)
    ├── output/                 # Final videos (.mp4) + metadata (.txt)
    ├── thumbnails/             # Video thumbnails (.png)
    └── prompts/                # Claude Code prompt templates

NotebookLM Automation

Automate NotebookLM interactions via Playwright browser automation.

Setup:

cd notebooklm-automator
mise install        # Install Python 3.13 + uv
mise run install    # Install deps + Playwright
mise run init       # Initialize Chrome profile
mise run login      # Login to Google (one-time)

Key features:

• Create notebook: Auto-create from episode number (finds PDF from status.json)

  mise run notebook -- 66

• Generate audio: Start podcast generation

  uv run notebooklm-automator audio generate \
    --notebook-url "https://notebooklm.google.com/notebook/xxx" \
    --language Polish

• Download audio: Download when ready (waits if still generating)

  uv run notebooklm-automator audio download \
    --notebook-url "https://notebooklm.google.com/notebook/xxx" \
    --output ./audio/15-glam.m4a

• Batch download: Download all ready audio from status.json

  uv run notebooklm-automator audio batch-download

• Generate slides: Create slides PDF from prompt

  uv run notebooklm-automator slides generate \
    --notebook-url "https://notebooklm.google.com/notebook/xxx" \
    --prompt-file ./prompts/slides.md \
    --output ./slides/15-glam.pdf

See notebooklm-automator/README.md for full documentation.

Full Workflow

0. Download Papers (Optional)

Batch download papers from curated lists:

# See template for input format
cat future/future-template.md

# Dry run to preview
mise run download -- future/my-papers.md --dry-run

# Download papers with auto-indexing
mise run download -- future/my-papers.md

Features:

• Auto-assigns episode numbers (fills gaps in status.json)
• Checks for duplicates by name/URL
• Tries multiple sources (arXiv, OpenReview, direct URL)
• Verifies PDF validity
• Updates status.json automatically
• Reports failed downloads for manual intervention

1. Add Audio from NotebookLM

Export podcast from NotebookLM and save to audio/:

audio/XX-paper-name.m4a

Naming convention: {number}-{paper-name}.m4a (e.g., 02-gpt.m4a)

2. Transcribe Audio

Run batch transcription (uses Whisper, Polish language):

mise run transcribe
# Or with custom parallelization:
mise run transcribe -- 4

Output: youtube/pl/transcripts/XX-paper-name.json

3. Generate Slides (NotebookLM)

Option A: Use existing prompt template

# Read transcript and create slide prompt
cat prompts/generate-slides-command.md

Option B: Use Claude Code

Run in fresh Claude Code session:

Read the transcript file at youtube/pl/transcripts/XX-paper-name.json
and create a detailed NotebookLM prompt for generating 10 presentation slides.

Then paste the generated prompt into NotebookLM to create slides PDF.

Save slides to: slides/XX-paper-name.pdf

4. Add Thumbnail

Generate thumbnail in NotebookLM or externally. Save to: youtube/thumbnails/XX-paper-name.png

Rename numeric thumbnails to match whitepaper names and compress:

mise run rename-thumbnails            # Rename + compress to <1.9MB
mise run rename-thumbnails --dry-run  # Preview changes

5. Generate Video

Prepare slides

mise run prepare -- 28  # Extracts PDF, normalizes images

Create concat.txt from slide timings

Use the template script which automatically handles intro/outro:

# Provide content slide durations (thumbnail/outro added automatically)
mise run generate-concat -- 28 --durations slide-01:180,slide-02:150,slide-03:120

# Preview without writing:
mise run generate-concat -- 28 --durations ... --dry-run

# From JSON file:
mise run generate-concat -- 28 --json timings.json

The script automatically adds:

• 5s thumbnail intro (duplicated to avoid ffmpeg drop bug)
• Your content slides with specified durations
• 5s silent outro with last-slide.png

Verify concat.txt

mise run verify-concat -- 28 --check-dims

Generate video

mise run video -- 28
# Or skip verification:
mise run video -- 28 --skip-verify

Or use Claude Code slash command

/generate-video 28

This runs the full video generation workflow with proper timings.

6. Output

Final files in output/:

• XX-paper-name.mp4 - Video file
• XX-paper-name-metadata.txt - Title, description, tags for YouTube

File Naming Convention

All files must follow: {XX}-{paper-name} where:

• XX = episode number with leading zeros (01, 02, ... 99, 100+)
• paper-name = lowercase, hyphenated paper identifier

Examples:

• 01-attention-is-all-you-need
• 73-raft
• 130-trusting-trust
• 200-pastry

Requirements

• whisper - Audio transcription (pip install openai-whisper)
• ffmpeg - Video generation (brew install ffmpeg)
• poppler - PDF to PNG (brew install poppler)
• imagemagick - Image processing (brew install imagemagick)
• jq - JSON processing (brew install jq)

Install Python dependencies:

mise run install

YouTube Upload Setup

Automated upload requires Google Cloud OAuth credentials.

1. Create Google Cloud Project

1. Go to Google Cloud Console
2. Create new project or select existing
3. Enable YouTube Data API v3:
   • APIs & Services → Library → search "YouTube Data API v3" → Enable

2. Create OAuth Credentials

1. APIs & Services → Credentials → Create Credentials → OAuth client ID
2. Application type: Desktop app
3. Download JSON → rename to client_secret.json
4. Move to .youtube-credentials/client_secret.json

3. Configure Playlists (optional)

Set playlist IDs via environment variables:

export YOUTUBE_PLAYLIST_LLM="PLxxxxxxxxxx"
export YOUTUBE_PLAYLIST_DISTRIBUTED_COMPUTING="PLyyyyyyyyyy"
export YOUTUBE_PLAYLIST_SECURITY="PLzzzzzzzzzz"
export YOUTUBE_PLAYLIST_NETWORKING="PLaaaaaaaaa"
export YOUTUBE_PLAYLIST_OPERATING_SYSTEMS="PLbbbbbbbbb"

Or create youtube/config.json:

{
  "playlists": {
    "llm": "PLxxxxxxxxxx",
    "distributed-computing": "PLyyyyyyyyyy",
    "security": "PLzzzzzzzzzz",
    "networking": "PLaaaaaaaaa",
    "operating-systems": "PLbbbbbbbbb"
  }
}

4. Upload Video

mise run upload -- 28           # Upload episode 28 (private)
mise run upload -- 28 --dry-run # Validate without uploading
mise run upload -- 28 --privacy unlisted

First run opens browser for Google OAuth. Token saved to .youtube-credentials/token.json.

Quick Start Example

# 1. Transcribe new audio
mise run transcribe

# 2. Prepare slides for episode
mise run prepare -- 28

# 3. Run Claude Code slash command (or generate video manually)
# /generate-video 28
mise run video -- 28

# 5. Upload to YouTube:
#    - Video: youtube/output/28-paper-name.mp4
#    - Metadata: youtube/output/28-paper-name-metadata.txt
#    - Thumbnail: youtube/thumbnails/28-paper-name.png