Video-Making: a video skill pack for AI coding agents

One installable agent skill (video-making), scripts and a tested pipeline that let an AI agent (Codex, Claude Code, or any agent that reads SKILL.md files) make narrated, captioned, music-backed videos entirely from code, with free local tools: Kokoro text-to-speech, word-exact karaoke subtitles, canvas animation captured one frame at a time by headless Chrome, ffmpeg mixing at -14 LUFS, and automatic quality checks.

Quick start

Linux (x86_64; Python 3.8+ bootstrap, Node 18+, ffmpeg with libx264):

git clone https://github.com/namra-030604/Video-Making && cd Video-Making && sh install.sh --install-system

macOS 14+ (Apple Silicon full video; Intel Mac automatically gets voice-only; Python 3.8+ or uv bootstrap):

git clone https://github.com/namra-030604/Video-Making && cd Video-Making && sh install.sh --install-system

Windows 10/11 x64 (PowerShell; Python 3.8+ bootstrap, Node and ffmpeg):

git clone https://github.com/namra-030604/Video-Making; cd Video-Making; powershell -ExecutionPolicy Bypass -File .\install.ps1 --install-system

The installer uses uv to create .venv with Python 3.13, downloads pinned models, fonts and Chrome, then links the skill. Install uv with brew install uv on macOS or winget install --id astral-sh.uv -e on Windows if the automatic pinned pip bootstrap is unavailable. Linux can use python3 -m pip install --user uv==0.13.7. Run python3 vsp.py install --check (or py -3 vsp.py install --check on Windows) first to see missing system tools and exact commands. --install-system (--apt) runs those commands only when no password prompt is needed.

For AI agents: Clone the repo and run the one install command above for the operator's OS. Then run python3 vsp.py verify and python3 vsp.py smoke on Linux/macOS, or py -3 vsp.py verify and py -3 vsp.py smoke on Windows; proceed with a project only when verify reports all required components PASS and smoke reports PASS. Create a project with python3 vsp.py new my-video and build it with python3 vsp.py build my-video on POSIX; use py -3 in place of python3 on Windows. The shell wrappers still work on POSIX; Windows uses vsp.py directly. See AGENT_INSTALL.md.

One skill: router -> master video -> components

video-making/
  SKILL.md                     the ONLY skill entry point: a router (keywords -> component -> what to run)
  components/
    master-video/MASTER.md     full end-to-end orchestration (default route for "make a video")
    script-writing/  voice-and-audio/  subtitles-karaoke/  vertical-9-16-framing/  short-form-editing/
    qc-and-render/  lessons-from-failures/  video-production-standard/      (each a COMPONENT.md)
    motion-studio/   COMPONENT.md + templates/ scripts/ references/        (music-led motion pieces)
    motion-graphics/ COMPONENT.md + fx.js + guides/ (18 effect guides) + gallery/ + test/

The router reads the request and sends the agent to one component: script/text, voice/music/SFX/loudness, motion graphics (or the maps guide), music-led promos, captions/karaoke, 9:16 framing, short-form editing, rendering/QC/resume, lessons, the production standard, or (default, for a whole video) the master component. python3 scripts/router_test.py checks the routing table against tests/router_cases.json: it is a deterministic keyword check, not an LLM test.

What you get

Folder	Contents
video-making/	the single skill (router SKILL.md, master component, 10 sub-components, 18 motion guides)
scripts/	the pipeline: make_voice.py, align_words.py, make_audio.py, stills.mjs, render.mjs, subtitle_qc.mjs, qc_final.py, framediff.py, build.sh, new_project.sh; checks: router_test.py, check_links.py
tests/router_cases.json	sample prompts and the component each must route to
lib/browser.mjs	a small dependency-free Chrome DevTools driver (no npm packages needed)
examples/hello-world/	a complete 540x960 narrated scene to copy
smoke-test/run.sh	end-to-end test of the whole pipeline
vsp.py, install.sh, install.ps1	portable installer, verification and pipeline; POSIX and PowerShell wrappers
requirements.lock.txt, requirements-voice.lock.txt	universal full-video and voice-only Python locks

Requirements

• Linux x86_64, macOS 14 or newer (Apple Silicon for full video; Intel Mac for voice-only), or Windows 10/11 x64. Intel Mac detection includes x86_64 processes under Rosetta. The voice-only tier uses onnxruntime 1.23.2 and skips Torch, MMS_FA, Chrome, and video rendering. Force that tier on any OS with --voice-only or VSP_VOICE_ONLY=1.
• Python 3.8+ or uv for the bootstrap; uv creates the runtime Python 3.13 venv. Python 3.12-3.13 is the supported runtime window. Without uv, install Python 3.13 or 3.12 (py -3.13/py -3.12 on Windows, python3.13/python3.12 elsewhere), then install uv. The default installer uses uv because it manages the interpreter.
• Full video needs Node 18+ and ffmpeg/ffprobe with libx264. Voice-only needs ffmpeg for audio mixing. install --check reports missing tools. VSP_FFMPEG, VSP_FFPROBE, VSP_NODE, VSP_BLENDER and CHROME override discovery. Chrome is otherwise downloaded as a pinned build.
• About 4 GB free disk and 2 GB free RAM. No GPU, API key or paid account. Set VSP_CACHE to move the cache; otherwise it is ${XDG_CACHE_HOME:-~/.cache}/video-skill-pack on every OS.

Download sizes (first install, cached afterwards)

Item	Size
Full-video Python venv (CPU PyTorch, onnxruntime, kokoro-onnx...)	~1.1 GB on disk; voice-only is smaller
Kokoro v1.0 int8 model + voices	~120 MB
MMS_FA alignment model (skipped in voice-only or with --skip-aligner)	~1.26 GB
chrome-headless-shell 156.0.8078.47 (skipped with --no-chrome or voice-only)	~126 MB zip, ~270 MB unpacked
Inter + Caveat fonts	~1.3 MB

Intel Mac and forced voice-only mode

sh install.sh --voice-only (or py -3 vsp.py install --voice-only on Windows) installs Kokoro, onnxruntime, the pinned core Python packages, voice models, and fonts. Intel Macs select this tier automatically, including an x86_64 Python process under Rosetta. Run python3 vsp.py new my-video, edit its project.json, then python3 vsp.py voice my-video and python3 vsp.py audio my-video. On Windows, replace python3 with py -3. The direct scripts/make_voice.py and audio mixing scripts also work. build, smoke, direct stills/render and subtitle QC stop with one explanation: these pins have no Intel Mac Torch video stack or supported Chrome rendering path. Use Apple Silicon macOS 14+ for the full video pipeline.

Important license notes

• The word-alignment model (MMS_FA) is CC-BY-NC 4.0: non-commercial use only. Captions in commercial videos need a different aligner (not included).
• Remotion (optional, --with-remotion) is free only for individuals and organisations of up to 3 people.
• Check the license of every music, SFX, image or footage file you add; the pack's own sounds are synthesized.

License and credits

The pack's own code and docs: MIT. Third-party tools and the projects that inspired parts of the pack are listed with their licenses in CREDITS.md and NOTICE. No third-party code is vendored except where CREDITS.md says so.