Animated Shorts Without an AI Video Generator: A Free Claude Skill That Draws Every Frame in Code

One of the fastest-rising creative projects on GitHub this week makes animated videos and contains no video model at all.

It is called claude-animated-shorts, it went public on November 1, and within two days its star count on shields.io passed 880. Its showcase piece is a 67-second vertical cartoon about a real event: the 1831 collapse of the Broughton bridge near Manchester, which reportedly gave way under a detachment of soldiers marching in step (the official inquiry blamed a badly made bolt). Bean-headed townsfolk with ink outlines, a painted 3D town, lens blur, a march on the soundtrack. It looks like the kind of explainer short that does well on every platform.

Every frame of it was written as code by an AI, then rendered on a normal PC.

That choice is the interesting part. While the big labs sell ever-better video generators, this project bets that for a whole category of creators, a picture you can edit line by line is worth more than a picture that looks real.

Why "edit" beats "generate" for explainers

Anyone who has made a short with an AI video tool knows the loop. You describe the shot, you get a clip, and something is wrong: the character's hand melts, the camera drifts left when you wanted right, the bridge has four towers in one shot and three in the next. You cannot fix the hand. You can only describe the shot again and hope.

For a cinematic mood piece, that lottery is tolerable. For an explainer, it is a real problem, because explainers live on precision. The arrow has to point at the right thing at the right word. The bridge has to stay the same bridge for a minute. A joke depends on a character reacting exactly as the narrator lands the line.

This skill sidesteps the lottery entirely. The characters are drawn by a small drawing library on a canvas. The town is assembled from free building props by Kenney, a well-known source of public-domain game art, recoloured to fit. Lighting, ink lines and blur are effects written in code. Camera moves are numbers in a shot list.

So when something looks wrong, there is a line to change. Late reaction? Shift it two frames. Wrong colour on the roofs? Change one value. That is the trade in one sentence: you give up variety and realism, and you get the ability to fix anything.

What the pipeline actually is

The skill is a set of instructions and templates for Claude Code, Anthropic's coding assistant, built on two pieces of free software. Remotion turns web code into video files. three.js draws 3D scenes in a browser. Put together, the AI writes a small web scene, and Remotion records it frame by frame.

The skill's main file lays out eleven steps. Here is the shape of them in plain terms.

Story first. You pick one surprising true story with a physical cause and effect you can show, plus an everyday comparison that explains the science. Facts and sources go at the top of the script file.

A script with a formula. About 3.7 words per second, 55 to 70 seconds long, in a fixed order: hook, setting, rising action, the moment it breaks, a deadpan double line, the explanation through the everyday comparison, the consequence, a callback twist. The rules say jokes must be researched facts and the narration should never describe what is on screen.

Voice, timed to the word. You record or generate a narration file, then make a file of word timings and fix them by hand. Everything downstream syncs to those timings, and each on-screen reaction is set to land zero to three frames before its word.

A shot plan. Shots, cuts, camera positions, impact shakes and focus changes all go into one file, story.ts.

World and characters. The AI builds the town, the sky, the lighting and the cast in code.

Sound. An event table drives synthesized sound effects and an original score that dips under the voice. The skill is firm that you lock the picture first, because any timing change afterwards means rebuilding the sound.

Check, render, check again. Test stills come first. Separate AI reviewers compare frames against the style references and return short verdicts. The final render runs through a quality script that hunts for black, white, frozen or popping frames and checks loudness.

According to the skill file, the full-quality render runs at about four frames per second, roughly ten minutes for the 67-second example. That is on whatever machine you own, with no per-clip bill.

Put this into practice

Start by watching. The repository's example is the fastest way to decide whether this house style suits your channel. If bean-headed cartoons in a painted town feel wrong for your subject, stop here and save yourself an afternoon.

If it fits, here is a realistic first run.

1. Install the parts. Copy the skill folder into ~/.claude/skills/cartoon-3d-short/. Install Node, Python 3, ffmpeg, and Remotion with its 3D add-on (@remotion/three). This is the slowest step if you have never set up a coding project.
2. Point Remotion's temp folder at a big drive. Every render bundle copies about 1.3 GB of assets into a temporary folder. The README tells you to set REMOTION_TMP to a large drive, and the skill file warns that old bundles pile up. Do this before your first render, not after your system drive fills.
3. Choose a story the formula can carry. It works best on one event with a clear physical chain: a bridge, a flood, a factory fire, an invention that failed. Abstract ideas without a visible cause and effect will fight the format.
4. Ask Claude Code for the short ("Use the cartoon short skill and make a short about…") and approve the script before anything else happens. The skill is built to wait for that approval. Use it. Script changes after the picture is locked are expensive.
5. Bring your own voice. The skill does not name a voice tool. Record yourself, or use a text-to-speech service you already pay for, and then spend real time on the word timings. Every gag depends on them.
6. Review stills before the full render. Stills render in seconds once the project is built. A ten-minute final render is a bad place to discover the roofs are teal.

Budget a full day for the first video, most of it setup and learning the shot file. The promise is that the second one goes faster, because the town, the characters and the sound engine already exist.

Where it breaks

One look, by design. The skill makes one style very well. A different look means rebuilding the world, the characters and the effects, which is a project, not a setting. If you want photoreal footage or varied art styles, this is the wrong tool.

Your voice is your problem. Narration is outside the skill. A flat voice will sink a well-animated short, and the hand-timing step is fiddly.

The reference frames are not yours to use. The repository includes a style study built from frames of existing shorts. The README says they are for style study only and that story, characters and voice should always be your own. The skill file goes further and says not to publish, upload or name the source. The code is MIT licensed; the reference images are not covered by that.

Remotion has a company threshold. Remotion is free for individuals and for teams of up to three people, including commercial use. According to Remotion's license FAQ, organizations of four or more need a paid company license, and if a client runs the project with you, both headcounts count. Solo creators are fine. A small studio should read the terms before building a pipeline on it.

It costs AI usage, not render fees. There is no per-clip charge, but the skill leans on Claude Code throughout, including separate reviewer agents for frame checks. The repository does not estimate that usage, so watch your plan's limits on the first run.

Disk and patience. Beyond the 1.3 GB temp copies, the setup assumes you are comfortable in a terminal. If "install Node" is a new sentence, expect a steep first evening.

"Viral" is the author's word. The repository's own description calls its shorts viral. Stars on GitHub tell you developers are curious, not that audiences will watch. Judge the example video, not the description.

The bigger idea

Generative video keeps getting better, and for many shots it will be the right tool. But this project points at a split that will matter more over time. Some work needs footage that looks real. A lot of explainer, education and history content needs something else: a consistent world, exact timing, and the ability to change one thing without changing everything.

For that second kind of work, an AI that writes the scene as code may simply be the better collaborator. It never forgets what the bridge looks like. It never invents a fourth tower. And when it gets a frame wrong, you can see exactly why.

Pick one true story you have always wanted to tell in under a minute. Watch the example, decide whether the style fits, and if it does, give the first one a full day. You will know by the end of it whether code-drawn animation belongs in your toolkit.

---

• Sources: claude-animated-shorts on GitHub · SKILL.md · README · Remotion license · Remotion license FAQ · Kenney assets