YouTube Publisher

Drive → YouTube → Transcript → Metadata → Cleanup.

Setup

1. Google OAuth (required for Drive download + 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 as google-oauth-client.json
4. Run the authorization flow to get google-oauth-tokens.json with scopes:
   • https://www.googleapis.com/auth/drive.readonly
   • https://www.googleapis.com/auth/youtube.upload
   • https://www.googleapis.com/auth/youtube

Place both files in one of these locations (checked in order):

• Skill root: youtube-publisher/google-oauth-client.json
• Config dir: youtube-publisher/config/google-oauth-client.json
• Workspace: ~/.openclaw/workspace/scripts/google-oauth-client.json
• Env vars: GOOGLE_OAUTH_CLIENT and GOOGLE_OAUTH_TOKENS (paths to files)

2. Transcription API key (at least one required)

Option A — Fireworks Whisper v3 Turbo (preferred: faster, cheaper, $0.0009/min):

• Set FIREWORKS_API_KEY env var, or
• Save key to youtube-publisher/config/fireworks-api-key.txt

Option B — Deepgram Nova-3 (fallback: good quality, smart_format):

• Set DEEPGRAM_API_KEY env var, or
• Save key to youtube-publisher/config/deepgram-api-key.txt

3. ffmpeg (required)

Must be on PATH. Install: apt install ffmpeg / brew install ffmpeg.

Procedure

Step 1: Extract file ID from Drive URL

Parse the Google Drive URL. Extract file ID from pattern /file/d/{ID}/. If user provides a raw file ID, use it directly.

Step 2: Download video from Google Drive

1. Refresh OAuth access token
2. Fetch file metadata (name, size, mimeType) via Drive API v3
3. Download file via ?alt=media endpoint to temp dir
4. Log file name and size

Step 3: Upload to YouTube

1. Create resumable upload session via YouTube Data API v3
2. Set initial metadata: placeholder title, privacy status from user (default: unlisted)
3. Upload video file
4. Save returned video_id
5. Log YouTube URL: https://youtube.com/watch?v={video_id}

Step 4: Extract audio

ffmpeg -i source.mp4 -vn -acodec libmp3lame -ab 64k -ar 16000 -ac 1 audio.mp3 -y

64kbps mono, 16kHz — optimized for speech transcription, minimal file size.

Step 5: Split into chunks

Split audio into 15-minute segments for parallel transcription:

ffmpeg -i audio.mp3 -ss {start} -t 900 -c copy chunk-{i}.mp3 -y

Step 6: Transcribe (Fireworks or Deepgram)

Fireworks Whisper v3 Turbo (default if FIREWORKS_API_KEY set):

POST https://audio-turbo.us-virginia-1.direct.fireworks.ai/v1/audio/transcriptions
model=whisper-v3-turbo, language={lang}, response_format=verbose_json

Deepgram Nova-3 (fallback if only DEEPGRAM_API_KEY set):

POST https://api.deepgram.com/v1/listen?model=nova-3&language={lang}&smart_format=true&paragraphs=true

Override with --transcriber fireworks|deepgram. Chunks transcribed in parallel (6 workers). Merge with time offset: chunk_index × 900 seconds.

Step 7: Generate timestamps

Extract topic boundaries from transcript. Rules:

• Minimum 3 minutes between timestamps
• First timestamp MUST be 0:00
• Format under 1 hour: MM:SS (e.g., 05:30, 45:12)
• Format 1 hour and over: H:MM:SS (e.g., 1:00:01, 1:25:30)
• NEVER write 60:01 or 75:30 — convert to 1:00:01, 1:15:30

Step 8: Update YouTube metadata

Update video via YouTube Data API v3 PUT /videos?part=snippet:

• Title: generate from transcript content (concise, descriptive)
• Description: summary + timestamps block + relevant links
• Tags: extract key topics
• Category: 28 (Science & Technology) unless user specifies otherwise
• Language: from user or auto-detect

Step 9: Clean up temp files

MANDATORY. No large files left behind.

rm -rf /tmp/yt-publisher-*/

Verify: find /tmp -maxdepth 1 \( -name "*.mp4" -o -name "*.mp3" \) -size +1M should return nothing.

Script

Full automated pipeline:

python3 scripts/publish.py "DRIVE_URL" [--privacy unlisted] [--language ru] [--transcriber auto|fireworks|deepgram]

Options:

• --privacy — public, unlisted (default), private
• --language — language code for transcription (default: ru)
• --transcriber — force backend: auto (default), fireworks, deepgram
• --skip-upload — transcribe only, don't upload to YouTube
• --video-id — update existing video metadata instead of uploading new
• --title — override auto-generated title
• --keep-files — don't delete temp media files after processing

Error Handling

• OAuth expired: refresh token automatically; if refresh fails, ask user to re-authorize
• Drive 404: file not shared or wrong ID — ask user to check sharing settings
• YouTube quota exceeded: daily upload limit is 6 videos — inform user, retry tomorrow
• Transcription timeout: retry failed chunk once; if still fails, skip and note gap
• Large files (>2GB): warn user about upload time; use chunked upload

Timestamp Format Reference

• 0 → 0:00
• 330 → 05:30
• 2712 → 45:12
• 3601 → 1:00:01
• 4530 → 1:15:30
• 7200 → 2:00:00