What It Does

Content Machine is a CLI tool that generates short-form videos (TikTok, Reels, Shorts) from a topic in 4 automated stages:

Pipeline Overview The 4-stage pipeline: topic → script → audio → visuals → video

Script Generation - AI writes a script with hooks, scenes, and visual directions
Audio Synthesis - Local TTS generates speech with word-level timestamps
Visual Matching - Stock footage is matched to each scene via keyword extraction
Video Rendering - Remotion composites everything with TikTok-style captions

# Generate a complete video from a topic
cm generate "5 productivity hacks for developers" --archetype listicle

# Or run stages independently
cm script "topic" --archetype howto
cm audio ./script.json --tts-engine kokoro
cm visuals ./audio-manifest.json --provider pexels
cm render ./visuals-manifest.json --template tiktok-captions

Architecture

content-machine/
├── src/
│   ├── cli/           # 24+ commands (generate, script, audio, etc.)
│   ├── pipelines/     # Stage orchestration
│   ├── llm/           # OpenAI, Anthropic, Google AI adapters
│   ├── audio/         # Kokoro TTS + Whisper ASR
│   ├── visuals/       # Pexels, gameplay clips
│   ├── render/        # Remotion compositions
│   └── schemas/       # Zod validation for all artifacts
├── remotion/          # Video templates and caption components
└── tests/             # 193 test files, 1400+ tests

Technical Highlights

Local Audio Pipeline (No API Costs)

TTS: Kokoro-js runs locally on quantized ONNX models (11 voices)
ASR: Whisper.cpp generates word-level timestamps
Reconciliation: Custom algorithm realigns Whisper output to original script for 4.4x better sync accuracy

Script Archetypes

Six built-in content formats:

listicle - “5 things you didn’t know about…”
versus - “X vs Y - which is better?”
howto - Step-by-step tutorials
myth - “Myth vs Reality”
story - Narrative format
hot-take - Opinion pieces

Video Templates

// Built-in Remotion templates
const templates = [
  'tiktok-captions',    // Word-highlighted captions
  'capcut',             // CapCut style
  'hormozi',            // Alex Hormozi style
  'karaoke',            // Karaoke-style highlights
  'gameplay-top',       // Split-screen with gameplay
  'gameplay-bottom',    // Gameplay on bottom half
];

Quality Gates

Every video is scored before export:

Caption sync accuracy (target: 99%+)
Audio quality rating
Visual coverage validation
Automated retry on quality failures

Research Integration

Optional research stage pulls from multiple sources:

cm research "AI agents" --sources hackernews,reddit,tavily --time-range week

Hacker News trending
Reddit discussions
Tavily web search
Brave Search API

Research findings are injected as evidence into scripts.

Current Status

v0.2.2 - Early development, actively maintained

Working:

Full end-to-end pipeline
All 1,400+ tests passing
NPM package published
Demo videos generated

Roadmap:

Review queue UI
Multi-platform scheduling
Analytics dashboard
Trend detection via MCP

Demo Videos

Split-Screen Gameplay + Content (Brainrot Template)

Split-screen gameplay demo Minecraft gameplay (top) + Subway Surfers (bottom) with AI-generated script overlay

Subway Surfers-Style with TikTok Captions

Subway Surfers captions demo Full-screen gameplay with word-highlighted TikTok-style captions

Latest News Listicle

Latest news listicle demo Auto-generated listicle with numbered badges, TikTok chunk captions, and Pexels stock footage

Why CLI-First?

Export-first design means:

Human review before publishing
CI/CD integration for batch generation
No platform API verification required
Full control over output quality

Content Machine