How to make 2.5D cartoon history shorts with Claude and Blender

A round-headed cartoon man with dot eyes stands in a throne room lit like a film set, candles flickering, dust in the light from the windows. A narrator tells you about a bronze jar from 132 AD that could tell which direction an earthquake came from. Sixty-five seconds later you know something you will repeat at dinner.

That is the 2.5D history Short: flat 2D characters inside full 3D sets, one voice, one true story. It is one of the clearest examples of a format where Claude can now do most of the production, because the hard parts are writing, staging and lighting, and Claude can drive Blender directly through MCP and look at what it rendered.

This tutorial is built mainly on Dumble's full course, credited below, plus the Blender MCP and animation tools it relies on. We have rewritten it as a method and added the reasoning behind each step.

Why this format works

Most AI animation dies in the first two seconds because it looks like every other AI animation. This format avoids the usual tells on purpose. The characters are deliberately simple, so nothing looks uncanny. The sets are deliberately rich, with warm light, real materials and depth of field, and that contrast between flat characters and lit worlds is the signature. Every shot has a reason. There is one human voice. And the story is true, so it can be checked, which earns trust no effect can buy.

As Dumble puts it in his course, "The animation is the delivery, not the point." Keep that sentence on the wall.

What creators published and claimed

Dumble (@Y0urDumble) published "How to build a 2.5D motion studio with Opus 5.5", a course built around a breakdown of a Ricks Clips short about Zhang Heng's seismoscope. He reports that the reference video reached 27 million views in eight days, which is a third-party figure he observed, and he says plainly that he does not know the channel's actual pipeline: his course is an open-source way to build the same kind of format, not a reveal of theirs.

In a companion post he lists the stack he used for a prototype: a brainstorming skill to pin down the story, a style library, Blender MCP for sets, a cartoon character kit with acted emotions, an animation principles skill, a tool that lets the agent listen to its own mix, and HyperFrames for the cut. A later post describes a clay-character variant with webcam motion capture and three hard-won rules about rigs and props.

The underlying idea that an agent can build, look at and fix a Blender scene is covered well in Rina's article on Blender MCP, which is written about another model but applies to any agent: expect a crude first blockout, and the value is in the inspect, adjust, re-render loop.

Step 1: break down a reference as a format

Pick a short in the style and analyse it as a format, not as content to copy. Extract frames at two per second, then have Claude write a reference file listing every cut with its timestamp, what is 2D and what is 3D in each shot, the camera move and its direction, the caption style, and where the hook and twist land.

Dumble's best touch is labelling every value as measured (read from a frame), estimated (by eye) or a decision (your creative choice). Claude's spatial estimates from images are approximate, so anything you build on should be checked. End the file with five decisions for your own original episode, and never copy characters, designs, script or story structure one to one.

What a breakdown of the reference found (his measurements)

FORMAT      1080x1920, 30 fps, 65 s
STORY       one true story from history, one narrator
CHARACTERS  flat 2D: round heads, dot eyes, thick outlines
SETS        full 3D: lit rooms, crowds, depth of field
CUTS        a new shot roughly every 2 s
CAMERA      slow push-ins and dollies, never random drift
CAPTIONS    1-2 words, bold white, centre-low
HOOK        a why / how / what-really-happened title

Step 2: house rules in CLAUDE.md

Claude Code reads CLAUDE.md on every run, so put the format there once and every episode starts from the same rules. The essentials: render at exactly 1080 by 1920 and never crop or upscale; flat characters that never get realistic shading; a new shot on every new idea, roughly every two seconds; camera moves with a reason; true stories only, with every claim sourced in a sources.md file; one motivated key light and warm practicals in every set; worn, varied materials. Then a banned list: centered titles on gradients, corner labels and timecodes, everything fading in, a whoosh on every cut, and any number on screen that is not in sources.md.

Step 3: a 160 word script with sources

A 65 second short read at a normal pace is about 150 to 170 words. That is the whole budget, and it forces good writing. Ask Claude for five true stories that fit a title pattern ("The object that did something surprising", "Why a thing always does X"), each with two sources you can open and a line on why both a 12 year old and a 50 year old would repeat it. Pick one, then ask for the script: one narrator, short spoken sentences, a hook that makes a promise in line one, a twist around the 40 second mark, and a last line that pays off the hook.

Then check the sources yourself. This is the one step you never hand off completely. A single wrong date in the comments costs more trust than any render can buy back.

Step 4: a shot list in frames, not seconds

Split the script into one shot per idea, each 1.5 to 3 seconds, and give every shot the same fields: which set and where in it, the framing and camera move with its reason, who is on screen with pose and expression, any hero prop, and a sound only if something physical happens. Reuse sets: a 65 second episode should need three to five, not thirty.

Then write the key beats in frame numbers. This is the habit that makes revisions fast, because "move the hit to frame 1260" is an instruction an agent cannot misread. Use physics as a sanity check: a ball falling about 30 centimetres takes roughly a quarter of a second, so seven or eight frames at 30 fps. Make it slower and it reads as floaty even to viewers who cannot say why.

Beat timing in frames (30 fps)

FRAME  TIME   EVENT
1215   40.5s  SHOT 18: wide on the jar, night, court asleep
1230   41.0s  push-in starts, scale 1.00 -> 1.06 over 60 frames
1248   41.6s  the west dragon's mouth opens
1254   41.8s  the ball drops
1262   42.1s  the ball hits the frog  (hit sound on this frame)
1266   42.2s  CUT to SHOT 19: the inventor's eyes open
1272   42.4s  caption "dropped" fully readable

Step 5: sets in Blender, in three passes

With Blender MCP connected, Claude writes Blender Python and can also look at the viewport, a camera or a render, which is the whole game: the agent checks its own work instead of guessing. Build each set in three passes and stop after each one for a look sheet.

• Block: grey shapes for walls, furniture and floor, with every camera from the shot list placed. Send a sheet of camera views.
• Light: one motivated key (the windows, a torch), warm practicals, an HDRI for fill. No three-point studio lighting. Send stills from three cameras.
• Dress: materials with wear and variation, small props, depth of field on every camera focused where the character will stand.

Poly Haven HDRIs and textures are CC0, so they can be used in a monetised video without credit. And make the hero prop, the thing the story is about, a proper 3D object with bevelled edges and wear in the crevices, checked on a six angle turntable before it goes into the set. It deserves its close-up.

Step 6: cut-out characters that sit in the world

The 2.5D look comes from flat images on planes inside the 3D set. In Blender, import character parts as image planes with an emission or shadeless shader so the set lights never shade them, and add a soft contact shadow under the feet so they stand on the floor instead of floating. That contact shadow is the detail most first attempts miss.

Design the character as separate SVG parts (head, body, arms, legs, costume) and an expression sheet: a handful of eyes and mouths that swap per beat, like old replacement animation. One outline weight, flat colour, no gradients.

Acting line for one shot

SHOT 07 (2.1 s)  "the whole court laughed"
  thought:      he knows he is right
  anticipation: eyes narrow first, 4 frames before anything moves
  action:       head turns to camera on the word "laughed"
  hold:         10 frames on the look so it reads
RULES  no idle bounce; every move starts with a look;
       swap mouths only on stressed words

Write an acting line like this for every shot before animating. The fastest way to spot generated 2D animation is a character that never thinks: it bounces and waves, but nothing happens in its face before it moves.

Step 7: your voice, word-timed captions and sound

Record the script yourself. A real voice with real breaths is the biggest trust signal in the format, and it keeps you clear of voice cloning questions. Run Whisper with word timestamps and build the caption layer from that file: one or two words at a time, cut exactly on each word's start, bold white with a dark outline, centre-low inside the safe area, with a pop only on stressed words.

For sound, write cues against the same frame numbers as the picture: the release, the hit, a moment where the music drops out. Use a sound only where something physical happens, never repeat a whoosh more than twice, and make the music resolve on the final frame. A soft lead-in on a sound file shifts its audible hit, so align the waveform peak to the frame, not the start of the file.

Step 8: review sheets and scoped fixes

Never watch a full export ten times. Have ffmpeg build a contact sheet at one frame per second and a phone test at 360 pixels wide, and hand them to Claude as a critic rather than the author. Ask specific questions: does every character have a contact shadow, is any character lit like a 3D object, does any shot run past three seconds without a new idea, are captions readable at phone size, is any claim on screen missing from sources.md.

Then give every correction four fields: where (the frame range), what is wrong, what changes, and what must stay fixed. A scoped edit stops the agent from wandering into shots that were already right. Blender can render a single frame or a range from the command line, so a fix never needs a full re-render.

A scoped revision

REVISE SHOT 18
WHERE    frames 1248-1266
PROBLEM  ball hits the frog at 1268, 6 frames after the sound
CHANGE   only the fall: start 1254, contact 1262, gravity ease-in
KEEP     the push-in, the cut at 1266, caption timing, all else
THEN     stills at 1254, 1258, 1262, 1266 and a 1215-1290 preview

Pitfalls from people who built it

• Characters lit like 3D objects. Keep them shadeless with a contact shadow.
• Rendering bigger and cropping. Render at 1080 by 1920.
• Too many sets. Three to five per episode is plenty.
• Idle bounce. Every move starts with a look.
• Unsourced facts. Keep sources.md next to the script and check it yourself.
• For the clay variant with motion capture, Dumble's rules: align rest poses before retargeting or arms pass through the body; grip props from the finger bones or objects float out of the hand; render close-ups of hands and feet before the full render.
• Letting an agent "clean up" a .blend freely. Save incremental versions and tell it exactly what it may change.
• Copying a channel's characters or scripts. Take the grammar, never the content.

Cost and time

Blender, the MCP server, Whisper and the animation skills are free and open source, and Poly Haven assets are CC0. The costs are Claude usage, render time on your machine and your review time. Dumble's advice for the first episode is to time every stage (story, sets, characters, animation, voice, sound, review); the second episode tells you what the first one really cost. Expect the first to be slow, because it is when you build reusable sets and the character rig.

On YouTube, original episodes with a real story, your voice and new sets are what keep a channel clear of the rules on mass-produced content. YouTube's AI disclosure is aimed at realistic altered content; clearly unrealistic animation is generally exempt, but label anything that could be mistaken for real footage.

Doing it in about 3 minutes with WhatToUpload Video Studio

The 2.5D Blender look is not a Video Studio style today. What is live: Kinetic type, Ink stickman and Diagram explainer, with more being added. For true-story history Shorts, Ink stickman is the closest: you type the story in one sentence and Claude writes the script and storyboard, our renderer draws every frame as code, ElevenLabs voices it and captions and sound are timed, in about three minutes. It is a fast way to find out which stories get watched before you build sets for them.

A 30 second video is 26 to 30 credits, and plans are $19, $49 and $99 a month.