I’ve watched a lot of AI video demos this year, and most of them blur together. Same glossy look, same strange hands, same eight-second clips stitched into something that almost tells a story. Then I saw this one and stopped scrolling.
A LinkedIn creator just released Havoc in Heaven (大闹天宫), a 21-minute shadow puppet film made with Claude Code and Claude Fable 5.2. It uses zero AI video models. Every frame is drawn in code. The story comes from the first seven chapters of Journey to the West, the 16th-century Chinese novel that gave us the Monkey King.
The author shared two minutes of the film, uncut, and it looks gorgeous. What really grabbed me, though, was how openly they talked about the effort. As the creator puts it, “One prompt is just a start.”
What It Actually Took
The original poster tracked the real cost of the film across their Claude Code sessions. These numbers are worth a slow read:
- 102 prompts: about 4,600 words of direction
- 59 hours: from the first prompt to the final version
- 22 rounds of review: with 20 versions rendered
- 20 changes of direction: some big, some small
I love this breakdown because it kills the “type one sentence, get a movie” myth. 4,600 words of direction is about the length of a long magazine feature. That’s directing, not prompting. The AI did the drawing, but a human made every creative call.
Why Shadow Puppetry, of All Things?
This part got to me. The expert points out that shadow puppetry is close to extinction. You can still learn it if you have the passion and the time, but for most people with busy schedules, that’s just not realistic.
So the author sees an AI-coded approach as a way to give this old art form new angles of storytelling. It doesn’t replace the craft. It gives the craft a new door that more people can walk through.
As large language models become more powerful, they give us new ways to tell stories.
That line is the heart of the whole project. Now let’s look at how the film actually got made.
The Workflow, Step by Step
The creator laid out a six-step process. I’ve added why each step matters, so you can borrow the logic for your own projects.
- Start with the story: an idea, the book, a script. A clear source keeps every later decision anchored, so the AI isn’t guessing what the film is about.
- Give it references: GPT Image 2.5 for characters and scenes. Visual references tell Claude what “good” looks like before it writes a single line of drawing code.
- Agree on a storyboard: six chapters, shot by shot. Locking the structure early means revisions fix details instead of tearing up the whole film.
- Let Claude Code draw everything in code, then render and edit the film: characters, scenes and motion all become code, so any frame can be changed precisely and re-rendered.
- Compose the music in code too: the author scored the film in their own AI Coded Music Studio, which keeps sound and picture in the same editable world.
- Review, critique, fix: in the author’s words, this loop is where the film is made. 22 review rounds and 20 renders prove it.
Step six is the one I keep coming back to. Most people quit after the first render looks “pretty good.” This innovator treated the first render as a rough draft.
The Tool Stack
Here’s everything the person who shared it used. It’s a surprisingly lean setup:
- Claude Code with Claude Fable 5.2: the director’s assistant that writes all the drawing and animation code
- JavaScript, headless Chrome and FFmpeg: the drawing engine, the invisible browser that renders frames, and the tool that turns those frames into video
- AI Coded Music Studio: the score
- GPT Image 2.5 in Codex: reference images for characters and scenes
If “headless Chrome” sounds scary, it just means a browser running in the background without a window. Claude draws each frame as a web page, Chrome captures it, and FFmpeg stitches thousands of those captures into a film.
How to Try This Yourself
You don’t need to start with 21 minutes. Here’s how I’d borrow this approach for a first experiment:
- Pick a tiny story: a fable, a myth or a single scene that runs 30 to 60 seconds.
- Choose a simple visual style: silhouettes, paper cutouts or flat shapes are perfect for code, which is exactly why shadow puppets work so well here.
- Write the storyboard before any code: list each shot, what moves and how long it lasts.
- Plan for many review rounds: keep notes on what feels off after each render and feed them back one change at a time.
A starter prompt to open your first Claude Code session might look like this: “Build a 30-second shadow puppet scene in JavaScript. A fox silhouette walks across a glowing paper screen, stops, and looks up at a moon. Render frames with headless Chrome and combine them into an MP4 with FFmpeg. Show me the storyboard first and wait for my approval before writing code.”
That last sentence matters. It copies the author’s habit of agreeing on structure before the AI starts drawing.
Why This Matters Beyond One Film
Most AI video today comes from diffusion models that generate pixels directly. They’re fast, but they’re hard to control. Characters drift, details flicker, and fixing one moment often means regenerating everything.
Code-based video flips that. Every element is defined, so it stays consistent. Every fix is targeted. And every version can be saved, compared and rolled back. For long-form storytelling, where a character has to look the same at minute one and minute twenty-one, that control is huge.
I think projects like this point to a new kind of creator: part writer, part director, part reviewer, with an AI coding partner doing the heavy lifting on execution.
The author shared the full 21-minute film in the comments, and it’s absolutely worth your time. Head over to the full LinkedIn post to watch it and see the clip for yourself!