scene-studio
This is a starting point for handing short-form video to AI, from planning to publish metadata. On the assumption that you work in Claude Code or Codex, two tracks (generative and remix), approval gates, license policy, and pipeline-performance policy are laid out ahead of time. That keeps AI from slipping on cuts, hooks, licensing, and speed.
This page is just a fork in the road. Use the table below to figure out where you should start, then jump straight there.
Where to start
Section titled “Where to start”| If you are | Go to |
|---|---|
| Not a full-time developer, new to the terminal, new to ffmpeg and yt-dlp | Beginner manual |
| Comfortable with code, but new to the two tracks, template filling, and the license gate | The beginner manual from prerequisites through first-run, then the expert manual |
| Already running a short-form pipeline | Expert manual, straight to architecture |
If you can’t decide which row you’re in, pick the first one. The beginner manual moves one action per step, so you can skim past anything already familiar.
How the two manuals differ
Section titled “How the two manuals differ”The expert manual reads at reference density. It assumes the vocabulary of the two tracks, approval gates, 5-structure template filling, license-policy L2, and pipeline-performance is familiar, and it leads with the reasoning behind decisions. Six sections: intro, architecture, tracks, verification, distribution, updates.
The beginner manual is step by step. It assumes you’re comfortable with the idea of AI making video for you, but that the terminal, ffmpeg, and yt-dlp might be new. With many external tools to install, the prerequisites are on the heavier side. It runs as an index plus eleven steps.
The two are the same base told at two paces. The tone differs, but the commands and workflow are identical.
Two tracks
Section titled “Two tracks”You run two tracks on one channel. Only the entry points fork; they reuse shared assets (caption styling, publish metadata, performance policy).
- Generative — originals, characters, new video. It flows from planning to scenes, images (Codex), and video (ffmpeg). The entry point is
scene-flow-orchestrator. - Remix — takes an existing video URL and flows through collection (yt-dlp), analysis (Whisper and hook detection), and auto-editing. The entry point is
remix-flow-orchestrator.
Ambiguous input (just “make a short-form video”) isn’t guessed; it stops and asks. Is there a URL? Is it a character and scenario?
Common to both
Section titled “Common to both”The first run is the spot where you call /start with 프로젝트 시작. /start verifies the environment (ffmpeg, OPENAI_API_KEY, per-track tools), asks which track you’ll use, and captures the channel identity (channel.json). It doesn’t install external tools automatically. Missing ones are only guided, and you install them yourself.
AI doesn’t make a finished product in one shot. Each phase stops at an approval gate, and a human approves, edits, or rolls back. Because outputs are kept as files in _workspace/ rather than chat, you come back to the exact spot even after a session break.
The remix track enforces a license policy. A source with no license label is refused, and it requires Fair Use of 30% or less and at least three transformation duties. Only a spot where the user explicitly accepts responsibility bypasses it.
The AI entry point is one line, /start, in Claude Code, and the repository root AGENTS.md in Codex. Both paths point at the same canonical document, AI_AUTOMATION.md.
It doesn’t publish automatically (S8). publish-copywriter makes only the publish metadata, and the platform upload is done by you.
The license is All Rights Reserved with a commercial license model. The payment flow and a lawyer-reviewed COMMERCIAL-LICENSE.md are still in progress, so both need to be finished before real sales. The manual and landing are public, so the page you’re reading right now is open as is.
Once you’ve picked your row, jump straight to the expert manual or the beginner manual.