Skate — AI-Powered YouTube → Viral Shorts (clipper) CLI

Turn long-form videos into viral-ready vertical shorts, entirely on your local machine. Free, no API keys, no cloud.

Saw a bunch of paid tools doing this — why pay when you can run it locally? Skate uses faster-whisper for transcription, Ollama for AI ranking, OpenCV for face tracking, and FFmpeg for rendering. Everything runs on your machine.

---

How It Works

Input (URL or file)
  → Download (yt-dlp)
  → Transcribe (faster-whisper)
  → Chunk transcript into segments
  → Score heuristically (no AI needed)
  → Rank with local LLM (Ollama)
  → Select best clips
  → Track faces for smart vertical crop
  → Render clips with subtitles burned in
  → Output organized shorts

---

Requirements

Tool	Purpose	Check
Bun	Runtime & package manager	bun --version
FFmpeg	Video cutting & processing	ffmpeg -version
yt-dlp	YouTube downloading	yt-dlp --version
Ollama	Local LLM for AI ranking	ollama --version
Python 3	Whisper & OpenCV scripts	python3 --version
faster-whisper	Local transcription	installed via setup-python
OpenCV	Face detection	installed via setup-python

Recommended Ollama Model

ollama pull llama3.2:3b

---

Installation

1. Clone and install dependencies

git clone https://github.com/yourusername/skate.git
cd skate
bun install

2. Set up Python environment (Whisper + OpenCV)

bun run setup-python

This creates a virtual environment at ~/.skate/venv and installs:

• faster-whisper — speech-to-text with word-level timestamps
• opencv-contrib-python — face detection via Haar cascades
• numpy — numerical processing

3. Link the CLI (optional)

bun link

Then you can run skate from anywhere.

---

CLI Usage

Process a local video

bun start -- clip video.mp4

Process a YouTube video

bun start -- youtube https://youtube.com/watch?v=abc123

Auto-detect URL or file

bun start -- https://youtube.com/watch?v=abc123
bun start -- video.mp4

Analyze only (skip rendering)

bun start -- analyze video.mp4

Render from cached analysis

bun start -- render video.mp4

Watch a directory for new files

bun start -- watch ./videos

Disable face tracking

bun start -- clip video.mp4 --no-crop
bun start -- youtube https://youtube.com/watch?v=abc123 --no-crop

By default, Skate tracks faces for smart vertical framing. Pass --no-crop to use a static center crop instead.

Check dependencies

bun start -- doctor

---

Configuration

Config is stored at ~/.skate/config.json and auto-created on first run.

{
  "model": "llama3.2:3b",
  "clips": 10,
  "minLength": 20,
  "maxLength": 90,
  "subtitleStyle": "minimal",
  "outputDir": "./output",
  "cacheDir": "~/.skate/cache",
  "ollamaUrl": "http://localhost:11434"
}

Options

Field	Default	Description
model	llama3.2:3b	Ollama model for ranking
clips	10	Number of clips to produce
minLength	20	Minimum clip length (seconds)
maxLength	90	Maximum clip length (seconds)
subtitleStyle	minimal	Subtitle style (minimal, tiktok, mrbeast)
outputDir	./output	Output directory
cacheDir	~/.skate/cache	Cache directory
ollamaUrl	http://localhost:11434	Ollama API URL

---

Output Structure

output/
└── <video-name>/
    ├── clips/
    │   ├── clip-01.mp4
    │   ├── clip-02.mp4
    │   └── clip-03.mp4
    ├── captions/
    │   ├── clip-01.srt
    │   └── clip-02.srt
    └── metadata.json

---

Project Structure

skate/
├── scripts/
│   ├── face_detect.py         # OpenCV face detection
│   ├── whisper_transcribe.py  # faster-whisper transcription
│   └── requirements.txt       # Python dependencies
├── src/
│   ├── commands/
│   │   ├── clip.ts            # Process local video
│   │   ├── analyze.ts         # Analysis only pipeline
│   │   ├── render.ts          # Render from cached analysis
│   │   ├── watch.ts           # Watch directory mode
│   │   └── doctor.ts          # Dependency checker
│   ├── core/
│   │   ├── pipeline.ts        # Main pipeline orchestrator
│   │   ├── downloader.ts      # yt-dlp integration
│   │   ├── transcriber.ts     # Whisper bridge
│   │   ├── chunker.ts         # Transcript chunking
│   │   ├── scorer.ts          # Heuristic scoring
│   │   ├── ranker.ts          # AI ranking bridge
│   │   ├── tracker.ts         # Face tracking
│   │   ├── renderer.ts        # FFmpeg rendering
│   │   └── subtitles.ts       # SRT/ASS generation
│   ├── ai/
│   │   ├── prompts.ts         # LLM prompt templates
│   │   ├── ollama.ts          # Ollama API client
│   │   └── ranking.ts         # AI ranking logic
│   ├── vision/
│   │   ├── face.ts            # Face detection
│   │   ├── scene.ts           # Scene detection
│   │   └── crop.ts            # Smart crop path
│   ├── ui/
│   │   └── tui.ts             # Terminal spinner UI
│   ├── config.ts              # Configuration loader
│   ├── types.ts               # TypeScript types
│   └── index.tsx              # CLI entry point
├── output/                    # Rendered clips
├── cache/                     # Cached downloads
├── models/                    # Local models
├── temp/                      # Working files
├── package.json
├── tsconfig.json
└── README.md

---

Pipeline Steps

Step	Description
Download	Pulls video from YouTube via yt-dlp (or uses local file)
Transcribe	Runs faster-whisper for speech-to-text with word-level timestamps
Chunk	Splits transcript into 30–90 second natural segments
Score	Heuristic scoring — speaking rate, emotion, story structure, hooks
Rank	Sends top candidates to Ollama for virality scoring
Select	Picks best clips based on combined heuristic + AI scores
Track Faces	Detects faces per frame via OpenCV for smart vertical crop
Render	Cuts clips, applies crop, burns in subtitles

---

npm Scripts

Script	Command
bun start	Run Skate
bun run dev	Run with watch mode (auto-restart on changes)
bun run typecheck	TypeScript type checking
bun run setup-python	Create venv and install Python deps

---

Caching

Skate caches aggressively at ~/.skate/cache:

• Downloaded video/audio files
• Transcripts
• Face tracking data
• Analysis results

Re-running is fast — only changed steps are re-executed.

---

Why Build This?

Every "AI shorts" tool out there charges $20–$50/month or requires API keys that bill per minute. Skate is:

• 100% local — nothing leaves your machine
• Free — no subscriptions, no API costs
• Private — your videos never hit a third-party server
• Customizable — swap models, tweak prompts, adjust scoring