@45ck/content-machine

Content Machine

Automated short-form video generation pipeline for TikTok, Reels, and Shorts.

Status: Early development. Not production-ready yet.

Demo videos

GitHub READMEs can’t reliably inline-play mp4, so this repo uses small previews (GIF/SVG) and links to full videos hosted elsewhere (GitHub Releases / YouTube / Vimeo).

• Demo hosting: https://github.com/45ck/content-machine/releases
• Guide: docs/guides/guide-demo-media-20260111.md

Demo 1: Split-screen gameplay + content (brainrot template)

Shows:

• Split-screen gameplay template with Minecraft (top) + Subway Surfers (bottom)
• cm import visuals to bring your own clips
• Gameplay slot selection (--gameplay-style subway-surfers)
• cm render --template brainrot-split-gameplay-top

Demo 2: Subway Surfers-style gameplay with TikTok captions

Shows:

• Full-screen Subway Surfers gameplay with word-highlighted captions
• cm import visuals + cm render --template tiktok-captions

Demo 3: Bring your own clips (no stock footage)

Shows:

• Minecraft clip imported via cm import visuals
• cm render --template tiktok-captions --caption-preset karaoke

Note: if you want real gameplay (Subway Surfers / Minecraft parkour), you must supply your own clips with the appropriate rights in ~/.cm/assets/gameplay/<style>/. This repo does not ship or download copyrighted gameplay footage by default. On Linux you can import clips with scripts/download-gameplay.sh (local files or URLs you have rights to).

Demo 4: Latest news listicle (TikTok captions + number badges)

Full clip: docs/demo/demo-4-latest-news.mp4

This demo uses:

• TikTok-style chunk captions (smaller default sizing)
• Auto list-number badges (#1, #2, …) with fade/scale animation
• List markers hidden in captions (so you don’t see 1: in the text)

Replicate the pipeline (exact commands):

# Node >= 20 recommended
nvm use

# (One-time) install deps + whisper for word-level timestamps
npm install
npm run cm -- setup whisper --model base

# 1) Research latest headlines (Jan 2026 example query)
npm run cm -- --yes research --no-angles \
  -q "Davos World Economic Forum Iraq US forces withdrawal North Korea official fired January 2026" \
  -s tavily -t week -l 12 \
  -o output/research.stories.json

# 2) Script (listicle)
npm run cm -- --yes script \
  --topic "global news this week" \
  --research output/research.stories.json \
  --duration 60 \
  -o output/script.stories.v2.json

# 3) Audio + timestamps (audio-first + reconcile for cleaner captions)
npm run cm -- --yes audio \
  -i output/script.stories.v2.json \
  -o output/audio.stories.v2.wav \
  --timestamps output/timestamps.stories.v2.json \
  --sync-strategy audio-first --reconcile --require-whisper --whisper-model base \
  --tts-speed 1.2 \
  --no-music --no-sfx --no-ambience

# 4) Visuals (Pexels)
npm run cm -- --yes visuals \
  -i output/timestamps.stories.v2.json \
  -o output/visuals.stories.v2.json \
  --provider pexels --orientation portrait

# 5) Render (TikTok preset + chunk mode)
npm run cm -- --yes render \
  -i output/visuals.stories.v2.json \
  --audio output/audio.stories.v2.wav \
  --timestamps output/timestamps.stories.v2.json \
  --caption-preset tiktok --caption-mode chunk \
  --caption-offset-ms -80 \
  --hook none \
  -o output/global-news.stories.v2.tiktok-smaller.badges-back.mp4

What is this?

Content Machine is a CLI-first pipeline that transforms a topic into a short-form video in four stages:

topic -> script.json -> audio.wav + timestamps.json -> visuals.json -> video.mp4

Terminology

This repo has a few loaded words (template, workflow, archetype, preset, pack).

• Canonical registry: registry/ubiquitous-language.yaml
• Generated glossary: docs/reference/GLOSSARY.md
• How it is enforced: docs/reference/ubiquitous-language-system-20260210.md

Run the full pipeline with cm generate or run each stage independently.

Ubiquitous Language (Important)

These words mean specific things in this repo:

• Archetype: script format (hook + structure + pacing rules). Used by cm script / cm generate.
• Template: render template (Remotion composition + render defaults). Used by cm render / cm generate.
• Workflow: pipeline orchestration preset (stage modes + defaults, optional exec hooks). Used by cm generate.

Full glossary: docs/reference/GLOSSARY.md.

Key features:

• LLM script generation into scene-based script.json
• Local TTS via kokoro-js with word-level timestamps from whisper.cpp
• Visual matching to scenes with stock footage providers (Pexels/Pixabay)
• Remotion rendering with word-highlighted captions
• Import Remotion templates as code templates (cm templates import ...) with explicit opt-in (--allow-template-code)
• CLI-first, stage-by-stage artifacts for iteration and CI
• Custom asset imports and workflow definitions for marketing pipelines

Pipeline

Stage	Command	Input	Output
script	cm script	topic string	script.json
audio	cm audio	script.json	audio.wav, timestamps.json
visuals	cm visuals	timestamps.json	visuals.json
render	cm render	visuals.json + audio.wav + timestamps.json	video.mp4

cm generate runs the full pipeline in order.

Quick Start

# Clone
git clone https://github.com/45ck/content-machine.git
cd content-machine

# Prereqs
# - Node.js >= 20 (recommended: use nvm + .nvmrc)
# - (optional) ffmpeg (for making README demo GIFs)

# Recommended Node setup (if you have nvm)
nvm install
nvm use

# Install dependencies
npm install

# Set up environment
cp .env.example .env
# Add your OPENAI_API_KEY

# Note:
# - In this repo: use `npm run cm -- <command>`
# - If installed globally: use `cm <command>`

# Run the full pipeline in dev mode
npm run cm -- generate "Redis vs PostgreSQL for caching" --archetype versus --output output/video.mp4 --keep-artifacts

Recommended Video Pipeline (Default)

For best audio/caption sync, the default pipeline runs audio-first with Whisper timestamps and reconciles ASR output back to the script text.

# One-time Whisper setup (needed for audio-first)
npm run cm -- setup whisper --model base

# Generate a short video with best-sync defaults
npm run cm -- generate "Redis vs PostgreSQL for caching" --archetype versus --output output/video.mp4 --keep-artifacts

Research (Latest News → Script → Video)

cm research gathers evidence across sources and writes research.json. You can then feed that into cm script / cm generate to keep scripts grounded in the latest info.

# 1) Research (no API keys required: Hacker News + Reddit)
npm run cm -- research -q "latest AI model news" -s hackernews,reddit -t day -o research.json

# 2) Generate a full video using that evidence
npm run cm -- generate "latest AI model news" --research research.json -a story --duration 35 -o output/news.mp4

Auto-research shortcut (runs Stage 0 automatically):

# Use `--research` with no value (or `--research true`).
# Always uses hackernews,reddit; also adds web if BRAVE_SEARCH_API_KEY is set,
# and adds tavily if TAVILY_API_KEY is set.
npm run cm -- generate "latest AI model news" --research -a story --duration 35 -o output/news.mp4

Custom Assets and Workflows

Bring your own audio + clips and still use cm captions/rendering:

# 1) Generate timestamps from existing audio
cm timestamps --audio assets/voiceover.wav --output output/timestamps.json

# 2) Build visuals.json from local footage
cm import visuals --timestamps output/timestamps.json --clips assets/clips --output output/visuals.json

# 3) Render with your assets
cm generate "Launch teaser" \
  --audio assets/voiceover.wav \
  --timestamps output/timestamps.json \
  --visuals output/visuals.json \
  --output output/video.mp4

Import Remotion Templates (Code Templates)

Import Remotion official/community templates (or any Remotion repo) into CM as a code template:

cm templates import https://www.remotion.dev/templates/tiktok --id remotion-tiktok

cm render -i visuals.json --audio audio.wav --timestamps timestamps.json \
  --template remotion-tiktok --allow-template-code -o out/video.mp4

Docs:

• docs/guides/guide-remotion-template-import-20260210.md
• docs/reference/cm-templates-reference-20260210.md

For repeatable pipelines, define a workflow in ./.cm/workflows/<id>/workflow.json:

cm workflows list
cm generate "Product recap" --workflow acme-custom --workflow-allow-exec --output output/video.mp4

Split-Screen Gameplay + Pexels (Brainrot Template)

This produces the same style as output/demo-subway-pexels/video.mp4: gameplay on one half, Pexels B-roll on the other, CapCut-style captions, muted gameplay audio.

Prereqs:

• .env has OPENAI_API_KEY (keyword extraction) and PEXELS_API_KEY (stock footage).
• Gameplay clips live in ~/.cm/assets/gameplay/<style>/ (or pass --gameplay <path>).

# 1) Prepare gameplay clips
mkdir -p ~/.cm/assets/gameplay/subway-surfers
cp /path/to/subway.mp4 ~/.cm/assets/gameplay/subway-surfers/

# 2) Generate script + audio + timestamps
cm script --topic "Redis vs PostgreSQL for caching" -o output/script.json
cm audio --input output/script.json --output output/audio.wav --timestamps output/timestamps.json --tts-speed 1.2

# 3) Match Pexels visuals + gameplay
cm visuals --input output/timestamps.json --provider pexels --orientation portrait \
  --gameplay-style subway-surfers \
  --output output/demo-subway-pexels/visuals.json

# 4) Render split-screen (gameplay on top, Pexels on bottom)
cm render --template brainrot-split-gameplay-top --split-layout gameplay-top \
  --input output/demo-subway-pexels/visuals.json \
  --timestamps output/timestamps.json \
  --audio output/audio.wav \
  --output output/demo-subway-pexels/video.mp4 \
  --download-assets

Layout options:

• Swap positions: --split-layout gameplay-bottom
• Force full-screen: --gameplay-position full or --content-position full

Troubleshooting:

• If Remotion fails to decode a gameplay clip, transcode to H.264 baseline: ffmpeg -y -i input.mp4 -vf "scale=1080:-2" -c:v libx264 -profile:v baseline -level 3.1 -pix_fmt yuv420p -preset veryfast -crf 23 -an -movflags +faststart output.mp4
• For detailed logs: --verbose plus REMOTION_LOG_LEVEL=verbose

Vendored Dependencies

This repo vendors several open-source projects for video generation:

vendor/
  remotion/              # React-based video composition
  short-video-maker/     # Reference patterns for Pexels + Kokoro
  open-deep-research/    # Deep research agent patterns
  ...

See VENDORING.md for details.

Configuration

# Required
OPENAI_API_KEY=sk-...

# Optional
PEXELS_API_KEY=...         # For stock footage
GEMINI_API_KEY=...         # For AI image generation (provider: nanobanana)
GOOGLE_API_KEY=...         # Alias for GEMINI_API_KEY (either works)
BRAVE_SEARCH_API_KEY=...   # For cm research --sources web
TAVILY_API_KEY=...         # For cm research --sources tavily (and auto-research in cm generate)
REDDIT_CLIENT_ID=...       # Optional (some Reddit features may use auth)
REDDIT_CLIENT_SECRET=...   # Optional

AI Images (Gemini)

Generate per-scene AI images as a first-class visuals provider (and animate them at render-time):

cm visuals --input timestamps.json --provider nanobanana --motion-strategy kenburns --output visuals.json
cm render --input visuals.json --audio audio.wav --timestamps timestamps.json --output video.mp4

Bring Your Own Visuals (Local)

Use your own folder of clips/images as a visuals provider:

# Use local video clips (best-match by filename/folders)
cm visuals --input timestamps.json --provider local --local-dir ./my-footage --output visuals.json

# Use local images (requires a motion strategy at render-time)
cm visuals --input timestamps.json --provider localimage --local-dir ./my-images --motion-strategy kenburns --output visuals.json

For deterministic mapping (no fuzzy matching), provide a manifest mapping sceneId -> assetPath:

cm visuals --input timestamps.json --provider pexels --local-manifest ./visuals-manifest.json --output visuals.json

Config Files

• Project config (auto-discovered by walking up from cwd): .content-machine.toml, content-machine.toml, .cmrc.json
• User config (optional): ~/.cm/config.toml, ~/.cm/config.json, ~/.cmrc.json
• Merge order: user first, then project (project wins)
• Override: CM_CONFIG=/abs/path/to/config.toml or cm --config /abs/path/to/config.toml ...
• Inspect: cm config paths, cm config show

Caption Fonts

Default captions use an Inter-first stack with Montserrat fallback. A small Inter font pack ships in assets/fonts/Inter for consistent rendering.

CLI overrides:

cm render --caption-font-family "Inter" \
  --caption-font-weight 700 \
  --caption-font-file "assets/fonts/Inter/Inter-Bold.woff2" \
  --input visuals.json --timestamps timestamps.json --audio audio.wav

Config defaults (.content-machine.toml):

[captions]
font_family = "Inter"
font_weight = 700
font_file = "assets/fonts/Inter/Inter-Bold.woff2"

JSON config supports multiple fonts:

{
  "captions": {
    "fontFamily": "Inter",
    "fontWeight": 700,
    "fonts": [
      {
        "family": "Inter",
        "src": "assets/fonts/Inter/Inter-Bold.woff2",
        "weight": 700
      }
    ]
  }
}

Export-First Design

Content Machine does not auto-publish to platforms. Instead, it generates a ZIP package:

output/
  2026-01-01-discord-bot-tutorial/
    video.mp4
    cover.jpg
    metadata.json
    upload-checklist.md    # Platform-specific instructions
    platform-hints/
      tiktok.md
      reels.md
      shorts.md

Why? TikTok/Instagram APIs require business verification and audit. Export-first means:

• Works immediately (no API approvals needed)
• Human reviews before publish
• Can customize per-platform before upload

Project Structure

src/
  cli/        # Commander.js entry points
  script/     # Script generation pipeline
  audio/      # TTS + ASR pipeline
  visuals/    # Footage matching
  render/     # Remotion integration
  core/       # Shared infrastructure
  prompts/    # Prompt templates
  hooks/      # Hook selection helpers
  package/    # Packaging outputs
  publish/    # Publish/export helpers
  research/   # Research command logic
  score/      # Scoring pipeline
  validate/   # Validation pipeline
  types/      # Shared types
  test/       # Test stubs/support
  index.ts    # Package exports

Roadmap

• Core CLI pipeline (script/audio/visuals/render)
• Remotion templates
• Playwright capture scenarios
• Review queue UI
• Analytics dashboard
• Trend MCP integration
• Multi-platform scheduling

License

MIT - See LICENSE

Contributing

PRs welcome. Please read:

• CONTRIBUTING.md
• CODE_OF_CONDUCT.md
• SECURITY.md
• SUPPORT.md

Credits

Built for Vibecord / Vibeforge.

Inspired by:

• Remotion - React video framework
• short-video-maker - Reference patterns
• open-deep-research - Research agent patterns