One request, one video, two approvals

One paragraph in, a finished video out

Picture typing this into Claude Code:

Make a 45-second animated explainer on how a language model picks its next word. The audience has never written code. Show me the script before you do anything else, and show me the scene plan before you build it.

A while later there is a video file in your folder with a script, a voice and cuts. Nobody opened an editing program. And twice along the way, the thing stopped and asked whether it was heading in the right direction.

That is OpenMontage, a free, open-source project that sets Claude Code up as a small production crew: a researcher, a scriptwriter, a narrator, an editor. You brief it the way you would brief a freelancer, and it works through the job in order.

What each stage hands you

The exact files vary. The order doesn't. For the request above, this is roughly what you would see.

First comes research. The agent works out what the video has to get across, what to leave out, and which facts it will lean on. You get a short brief, mostly for its own use.

Then the script: about a hundred words of narration, shaped to fit 45 seconds. This is the first stop. You read it like any draft. Maybe the opening is slow, or the second analogy muddies the first. You say so in a sentence and it rewrites.

Then the scene plan: a list of scenes, what is on screen in each, which line of narration it carries, how long it stays. Second stop. Scene four looks like generic stock footage? Tell it to make that one a diagram. That costs you a sentence too.

Only now does the expensive part begin. The voice is recorded, visuals are gathered, everything is cut to the narration and rendered. Skip the stops and a wrong assumption in the script would surface here, in a finished file, with the whole pipeline to run again.

What you could make with it

A product explainer for the landing page. A lesson for a course, where the script is the thing you already know how to write. A short clip for social. And the one I would try first: a weekly update video for your team. The structure repeats every week, so after two rounds of approvals you know exactly what to look for at each stop.

If you are curious why an AI can make video at all, I wrote about that in A video is a function of time.

Why it stops, and why you should copy that

A wrong assumption in step one doesn't stay in step one. It gets built on through steps two to six, and you find it in the last file. Fixing it then means redoing everything.

The pipeline moves your attention to where a correction is cheapest. At the script, a change costs one sentence. At the scene plan, one line in a list. After the render, it costs the render.

The same shape works far from video. Say you ask an agent for a monthly board pack: pull the numbers, draft the commentary, build the slides. Put the stop after the commentary outline. If it misread which number matters, you find out before it has built twenty slides around it.

Who this is for, and what it costs

This one is for people comfortable in a terminal. You need Python 3.10 or later, FFmpeg, Node.js 18 or later, and an AI coding assistant such as Claude Code, which needs a paid Claude plan. Setup is three commands:

git clone https://github.com/calesthio/OpenMontage.git
cd OpenMontage
make setup

Then you open the folder in Claude Code and describe the video. If you want to see something before committing, make demo renders sample videos that need no keys.

Basic use needs no API keys at all. Narration comes from Piper, and footage from free sources such as Archive.org, NASA and Wikimedia Commons. Expect it to look like it: free voices and archive footage are weaker than what paid providers sell. You can add optional paid keys for more tools. The project's own README estimates roughly $0.15 to $1.50 per prompt that uses paid providers. That is their figure, not mine.

One more thing. The license is AGPL v3. If you plan to build a product on top of it, read the license before you start, not after.

What to do this week

You don't have to install anything to use the idea. Take the next multi-step job you give an agent and find the one stage that is slow or expensive to redo. Tell the agent to stop right before it and show you the work. Then read it properly.