🎬 YouTube Publisher

🇷🇺 Читать на русском

An AI agent skill that turns "upload this recording to YouTube" into a fully automated pipeline.
Works with Claude Code, Codex, Gemini CLI, OpenClaw, or any agent with terminal access.
Your agent downloads from Drive, uploads to YouTube, transcribes, generates timestamps and metadata — you just say the word.

---

The Problem

You recorded a meeting, a lecture, a workshop. It's sitting in Google Drive. Now what?

You download a 2GB file, drag it into YouTube Studio, wait for the upload, try to come up with a decent title, scrub through an hour of video to place timestamps, write a description, copy-paste everything, publish, delete the local file. That's 30-60 minutes of mind-numbing work per recording — and you have dozens of them every month.

The Solution

Tell your agent:

"Upload last Friday's recording to YouTube"

The skill handles the entire pipeline:

• ⬇️ Downloads from Google Drive (no local file juggling)
• 📤 Uploads to YouTube (resumable, handles large files)
• 🎤 Transcribes audio via Fireworks Whisper or Deepgram Nova-3
• ⏱️ Generates timestamps from topic boundaries
• 📝 Creates title & description from transcript content
• 🧹 Cleans up all temp files automatically

~5 minutes for a 1-hour video. Zero manual work. The agent does everything.

Install

With Any AI Agent

The skill is two things: SKILL.md (instructions) and scripts/ (code). Copy them into your agent's workspace:

git clone https://github.com/smixs/youtube-publisher.git

# Claude Code / Codex / Gemini CLI — into your project
mkdir -p skills/youtube-publisher
cp youtube-publisher/SKILL.md skills/youtube-publisher/
cp -r youtube-publisher/scripts skills/youtube-publisher/

# OpenClaw
cp -r youtube-publisher/{SKILL.md,scripts} ~/.openclaw/workspace/skills/youtube-publisher/

That's it — SKILL.md + scripts/. Everything else in this repo (README, LICENSE, .github) is packaging for GitHub, not part of the skill.

The agent reads SKILL.md, understands the pipeline, and runs it. Just say:

• "залей запись созвона на ютуб"
• "upload this Drive recording"
• "транскрибируй и опубликуй"

As a Standalone CLI

The skill includes a Python script that works independently:

python3 scripts/publish.py "https://drive.google.com/file/d/abc123/view"

Setup

1. Google OAuth (required)

You need OAuth credentials for Google Drive (download) and YouTube (upload):

1. Go to Google Cloud Console
2. Create a project → enable Google Drive API and YouTube Data API v3
3. Create OAuth 2.0 credentials (Desktop app) → download JSON
4. Rename to google-oauth-client.json, place in the skill's root or scripts/ folder
5. Run python3 scripts/setup_oauth.py to authorize (saves tokens next to client file)

Required scopes: drive.readonly, youtube.upload, youtube

2. Transcription API Key (at least one)

Fireworks AI (recommended: $0.0009/min — a 1-hour video costs 5 cents)

export FIREWORKS_API_KEY=your_key
# or save to skills/youtube-publisher/fireworks-api-key.txt

Deepgram Nova-3 (alternative: $0.0077/min, great quality)

export DEEPGRAM_API_KEY=your_key
# or save to skills/youtube-publisher/deepgram-api-key.txt

3. ffmpeg

sudo apt install ffmpeg   # Ubuntu/Debian
brew install ffmpeg        # macOS

How the Skill Works

The agent reads SKILL.md and follows the pipeline step by step:

Google Drive → Download → Upload to YouTube → Extract audio →
Split into 15-min chunks → Transcribe in parallel (6 workers) →
Merge transcript → Generate timestamps → Update YouTube metadata →
Clean up temp files

Key Details

• Audio extracted at 64kbps mono, 16kHz (optimized for speech, ~1MB/min)
• Chunks transcribed in parallel, merged with time offsets
• Timestamps placed at topic boundaries (minimum 3-min gaps)
• YouTube auto-links timestamps as clickable chapters

CLI Options

Flag	Default	Description
--privacy	unlisted	public, unlisted, or private
--language	ru	Language code for transcription
--transcriber	auto	auto, fireworks, or deepgram
--skip-upload	—	Only transcribe, don't upload
--video-id	—	Update existing video metadata
--title	—	Override auto-generated title

Credential Lookup Order

The script searches for credentials in this order:

1. Skill root → skills/youtube-publisher/
2. Scripts dir → skills/youtube-publisher/scripts/
3. Workspace → ~/.openclaw/workspace/scripts/
4. Environment variables

Requirements

• Python 3.8+ (stdlib only, no pip packages)
• ffmpeg on PATH
• Google Cloud project with Drive + YouTube APIs
• Transcription key (Fireworks or Deepgram)

Limitations

• Source must be in Google Drive (no local file upload yet)
• YouTube daily upload quota: 6 videos
• Timestamp generation is heuristic (works best for structured content)
• Title/description generation is basic — agent review recommended

License

MIT

Author

Serge Shima · TDI Group · AI Masters