🎬 Video Studio — a film studio you drive from chat

Turn one idea into a finished, on-brand film — in a few hours, from a chat prompt. A Claude Code skill that runs the whole pipeline for you: brand & storyboard intake → capture or generate → edit, grade & localise → deliver. No timeline app. No agency. No shoot day.

---

What is this?

Video Studio is a Claude Code skill: a SKILL.md + reference playbooks + scripts that teach Claude to produce and edit video end‑to‑end. You describe the film you want in plain language; the skill orchestrates the entire pipeline and only stops for two sign‑offs (the storyboard, and final approval).

It does two things, and combines them:

• 🎞️ Edit existing footage — inventory clips (local files or URLs), auto‑select the right shots by subject (it reads the frames, not just filenames), then trim, grade and cut to a polished film.
• 🤖 Full‑AI — keyframe‑first generation: build a consistent start‑frame per shot, animate it with image‑to‑video, letter it, score it, and cut it — never blind text‑to‑video.

…held to one premium quality bar the whole way. This isn't a meme‑clip generator; it's built to produce work a brand would pay for.

Why you'll want it

• From chat. No editor, no render farm, no manual export. Prompt → film.
• On‑brand by default. It adopts your existing style guide (or creates one), builds a reusable leader + outro bracket, and holds the lines you set on every film.
• A real quality bar. Crafted transitions (no hard‑cut‑by‑default), heavy readable titles with backing, one consistent grade + grain, a hook in the first 1.5s, audio levelled & ducked. It refuses to ship "technically works."
• Honest & scriptable. Proven tools glued into one pipeline — ffmpeg, ImageMagick, nano-banana, fal.ai through one small library (scripts/lib/fal.sh: key, upload to fal's own storage, submit/wait/result, a cost log per production). No black boxes. It even improves its own playbook after every job.
• Self‑checking. A tooling preflight runs before every job and stops and asks rather than silently degrading. A one‑time doctor tells you if your machine is ready. And a measurable QA gate (scripts/qa-check.sh) checks every deliverable — colour tags, size, loudness, rhythm, dead cuts, AI label — before anyone looks at a frame strip.

⚡ Quickstart

1. Install it — as a plugin (recommended) or a plain skill.

In Claude Code:

/plugin marketplace add RichardTheuws/video-studio
/plugin install video-studio@theuws-video

Or manually, as a user‑level skill:

git clone https://github.com/RichardTheuws/video-studio.git
bash video-studio/install.sh    # copies the skill into ~/.claude/skills/video-studio/

2. Check your machine is ready:

bash ~/.claude/skills/video-studio/scripts/doctor.sh

The doctor reports exactly what you have, what's missing, and which mode you can run today (footage / AI / full) — with copy‑paste install commands for anything absent.

3. Add your keys (only what your mode needs) to ~/.config/video-studio/.env:

FAL_AI_KEY=...            # image-to-video + AI stills (full-AI mode)
GEMINI_API_KEY=...        # nano-banana keyframes (full-AI mode)
ELEVENLABS_API_KEY=...    # optional: voices / SFX

4. Just ask, in chat:

"Make a 60‑second launch film for our new feature. Adopt our brand from brand/style.md. Vertical cutdown too."

Claude loads the skill, runs the preflight, walks the intake gate, shows you a storyboard to approve, then produces the film. That's it.

New to Claude Code? It's Anthropic's agentic CLI — install it, cd into any project, and the skill activates automatically on any video task. See claude.com/claude-code.

🩺 Requirements

The doctor (skills/video-studio/scripts/doctor.sh) checks all of this for you. In short:

	Tool	For	Install
Core	ffmpeg · jq · curl	montage, transcode, grading, audio	brew install ffmpeg jq
Full‑AI	nano-banana (Bun + Gemini)	consistent keyframes / stills	see reference/tooling.md
	FAL_AI_KEY	image‑to‑video (Seedance 1.5 Pro / Kling v3 Pro)	fal.ai
Titles	node ≥ 22 (HyperFrames via npx, pinned)	the title layer: one composition for 9:16 + 16:9	brew install node
	ImageMagick + pango	text in the world (perspective on a surface) + fallback titles	brew install imagemagick pango
Footage	yt-dlp	only to pull clips from URLs	brew install yt-dlp

Primary platform: macOS. Most of it is portable to Linux (swap brew for your package manager).

How it works

  preflight ─► INTAKE (gate) ─► STORYBOARD (your sign-off) ─► capture / generate
                                                                     │
        deliver + docs ◄─ localise ◄─ edit, grade & score ◄─────────┘

The intake gate is where most of the time is saved: before a single frame is produced, the studio settles the brief and the brand — does a style guide exist to adopt, or must one be created? Is there a leader/outro? Does this need a storyboard? Which shots need generated keyframes? Getting this wrong is the most expensive mistake in any pipeline, so it's an explicit, non‑negotiable step.

Deep‑dive playbooks live in reference/:

This is the beginning of the README. Read the rest at the original. Read the full README ↗