Viral Clipper

Viral Clipper Skill

Build an end-to-end autonomous pipeline that:

1. Discovers viral-potential YouTube videos in defined niches (daily, no manual input)
2. Downloads, transcribes via faster-whisper on GPU, finds best moments via Gemini
3. Renders 9:16 vertical clips with karaoke captions using FFmpeg + NVENC
4. Auto-posts to YouTube Shorts via official API, with safe rate-limiting
5. Routes different niches to different YouTube channels

Architecture (split-brain)

LAPTOP (GPU)                    VPS (always-on)              Browser
┌────────────────┐              ┌──────────────────┐         ┌───────────────┐
│ worker.py      │  long-poll   │ clips-api        │  HTTPS  │ /clips page   │
│  yt-dlp        │ ◀─────────── │  FastAPI :5903   │ ◀────── │ submit + view │
│  faster-whisper│              │  SQLite queue    │         │ status table  │
│  Gemini        │              │  approval store  │         └───────────────┘
│  NVENC encode  │ ─────────▶   │                  │
│                │  upload MP4  │ clips-poster     │
│ discovery.py   │              │  YouTube API     │ ──────▶ Channel A
│  yt-dlp search │ ─────────▶   │  niche routing   │ ──────▶ Channel B
│  RSS channels  │  POST job    │  rate throttle   │ ──────▶ Channel N
└────────────────┘              └──────────────────┘

Why split: YouTube bot-detects datacenter IPs (VPS yt-dlp gets blocked). Laptop residential IP works. But VPS must be always-on for queue + posting (laptop sleeps).

When to use

Trigger this skill when the user asks for any of:

• "Build me a viral clip pipeline / OpusClip clone"
• "Auto-post YouTube Shorts from long videos"
• "AI clip selection from podcasts / interviews"
• "Multi-channel YouTube automation"
• "Klap.app / vidiq alternative"
• "Auto-generate shorts from YouTube channels I follow"

When NOT to use

• Single one-off video editing → use video-edit skill instead
• TikTok-only / Instagram-only → user needs different posting modules
• No GPU available on user's machine → suggest cloud GPU (RunPod) instead

Prerequisites checklist (verify before building)

1. Laptop with NVIDIA GPU (RTX 30/40/50 series, 8GB+ VRAM)
2. Linux VPS with sudo + nginx (Ubuntu 22+ recommended)
3. User has a domain pointing to VPS (for HTTPS subsite + OAuth callbacks)
4. Free Gemini API key — get at https://aistudio.google.com/app/apikey
5. Per channel: Google Cloud project + YouTube Data API v3 enabled + OAuth Desktop client
6. Python 3.11+ on both laptop and VPS
7. FFmpeg with NVENC support on laptop
8. Node.js 18+ on laptop (for yt-dlp js-runtime)

Critical gotchas (DON'T MAKE US REDISCOVER THESE)

These bit me hard during development. Encode them into your build:

1. yt-dlp on VPS gets bot-flagged

YouTube returns "Sign in to confirm you're not a bot" for any datacenter IP. Solutions in order of preference:

• Run yt-dlp on the laptop, not VPS (this skill's chosen approach)
• Cookies file from a real browser session (--cookies cookies.txt)
• Residential proxy ($$$)

2. Windows scheduled tasks /SC ONLOGON requires admin

For autostart without admin, use the Startup folder instead:

• C:\Users\<USER>\AppData\Roaming\Microsoft\Windows\Start Menu\Programs\Startup\
• Drop a .vbs that launches Python silently (window mode 0)

3. YouTube OAuth in "Testing" mode = 7-day refresh tokens

Refresh tokens auto-expire weekly. Either:

• Submit app for verification (1-4 weeks, free)
• Re-run OAuth weekly (annoying but free, no review)

4. YouTube Data API quota = 6 uploads/day default

Each videos.insert = 1,600 quota units, daily limit = 10,000. So 6 uploads max. Bake throttling in:

• 2h between posts
• max 3 posts/day per channel for new channels (avoid spam flag)
• Request quota bump via Google form to scale

5. PowerShell em-dashes break script silently

Em-dashes (—), curly quotes ("") in PowerShell scripts cause ParameterBindingArgumentTransformationException. Use ASCII only.

6. Cloudflare caches HTML ~5 min after deploy

Tell user to hard-refresh (Ctrl+Shift+R) or wait 5 min after each npm run build.

7. yt-dlp needs --js-runtimes node:node in 2026

YouTube changed their player to require JS evaluation. Without node, all extractions fail with cryptic errors.

8. Pillow 11.2.1 fails on Python 3.14 from-source

Use prebuilt wheels: pip install --only-binary :all: pillow

9. ytsearch{N}: works, ytsearchdate{N}: does NOT

The ytsearchdate URL scheme isn't recognized — it's just ytsearch5:query for top 5 results.

10. Channel @handles can collide

@allin is the Chamath/Sacks/Friedberg podcast. @allinpodcast is a D&D show. Always verify with yt-dlp --get-title "@HANDLE/videos" before adding to discovery config.

11. SQLite schema migrations need lazy-add

Don't drop tables, ALTER on init:

cols = [r[1] for r in c.execute("PRAGMA table_info(table)").fetchall()]
if "newcol" not in cols:
    c.execute("ALTER TABLE table ADD COLUMN newcol TEXT")

12. nginx proxy must be added to the HTTPS server block

Look for the existing 443 server block (often in a separate file like /etc/nginx/sites-available/main). Adding to the HTTP-only :80 block will return 404 because traffic comes via HTTPS.

Build sequence

Always follow this order — earlier steps unblock later ones:

Phase 1: Verify prerequisites

• nvidia-smi to confirm GPU + NVENC
• ffmpeg -encoders | grep nvenc to confirm h264_nvenc
• python -c "from faster_whisper import WhisperModel" to confirm faster-whisper
• node --version to confirm Node 18+
• SSH to VPS works
• Domain DNS resolves to VPS

Phase 2: VPS API

1. Copy templates/vps-api/* to /opt/clips-api/ on VPS
2. Replace YOUR-DOMAIN.com with user's domain
3. python3 -m venv venv && ./venv/bin/pip install -r requirements.txt
4. Generate two random tokens:

   WORKER_TOKEN=$(python -c "import secrets; print(secrets.token_urlsafe(32))")
   ADMIN_TOKEN=$(python -c "import secrets; print(secrets.token_urlsafe(32))")
   

5. Write /etc/clips-api.env with CLIPS_DATA_DIR, CLIPS_WORKER_TOKEN, CLIPS_ADMIN_TOKEN
6. Install + enable clips-api.service and clips-poster.service
7. Add nginx reverse-proxy block to existing HTTPS server:

   location /clips-api/ {
       proxy_pass http://127.0.0.1:5903/;
       proxy_set_header Host $host;
       proxy_set_header X-Real-IP $remote_addr;
       proxy_read_timeout 90s;
       client_max_body_size 500M;
   }
   

8. Verify: curl https://USER-DOMAIN/clips-api/health returns {"ok":true}

Phase 3: Laptop worker

1. Create C:\Users\<USER>\clips-pipeline\ (Windows) or ~/clips-pipeline/ (Mac/Linux)
2. Copy templates/worker/* into a worker/ subdir
3. Edit config.py:
   • Set VPS_API_BASE to https://USER-DOMAIN/clips-api
   • Set WORKER_TOKEN to the value generated above
   • Add Gemini API keys (1+ keys, more = better rate-limit handling)
4. Install Python deps:

   pip install yt-dlp faster-whisper google-genai requests
   

5. Test single video:

   python worker.py --once "https://www.youtube.com/watch?v=Ks-_Mh1QhMc"
   

   Should produce 6 MP4s in output/test_TIMESTAMP/

Phase 4: Discovery configs

1. Copy templates/discovery_configs/*.json to worker/discovery_configs/
2. Edit per niche — remove _comment, customize:
   • queries[] — keyword searches (yt-dlp ytsearch5:query)
   • channels[] — @handles or UCxxxx channel IDs
   • filters — duration/age/views/language minimums
   • limits — max videos per run, max pending jobs

Phase 5: YouTube OAuth (per channel)

For EACH channel:

1. Open https://console.cloud.google.com signed in as the burner account that owns the channel
2. Create new project named clips-uploader-NICHE
3. Enable YouTube Data API v3
4. Configure OAuth consent screen (External, scopes: youtube.upload + youtube.readonly, add the burner email as test user)
5. Create OAuth Client ID, type Desktop app, download JSON
6. On laptop, run OAuth helper (browser opens, sign in with burner account):

   python scripts/run_oauth.py
   

7. SCP token to VPS: /etc/clips-tokens/youtube_NICHE.json
8. chmod 600 /etc/clips-tokens/youtube_NICHE.json

Phase 6: Web UI (optional but recommended)

1. Copy templates/web/page.tsx to user's Next.js project at app/clips/page.tsx
2. Build + deploy site
3. Sign in at https://USER-DOMAIN/clips/ with admin token
4. Toggle "Auto-approve" if you want zero-touch posting

Phase 7: Autostart

1. Windows worker autostart: copy scripts/ClipsWorker.vbs to %APPDATA%\Microsoft\Windows\Start Menu\Programs\Startup\
2. Discovery daily run: create scheduled task via scripts/install_tasks.bat
3. VPS services: already enabled via systemd — systemctl enable clips-api clips-poster

Phase 8: Test E2E

1. Submit one test job via UI (single YouTube URL)
2. Watch worker logs, confirm clips render
3. Manually approve one clip, confirm it posts to YouTube
4. Check the channel's "Your videos" tab — should see the upload

File reference

Customization knobs

When user asks to tweak behavior, these are the levers:

Adding a new niche later

1. Drop worker/discovery_configs/{newniche}.json (copy an existing one as template)
2. Run OAuth flow signed into the new YouTube account
3. SCP token: /etc/clips-tokens/youtube_{newniche}.json on VPS
4. Discovery and posting auto-pick it up — no code changes

See also

• docs/SETUP.md — full step-by-step setup guide
• docs/TROUBLESHOOTING.md — common errors and fixes