How to Build Stop-Motion Motion Graphics Inside Claude (For About a Dollar a Scene)

TL;DR: You can generate paper-cutout, claymation, and anime-style stop-motion animations without After Effects or an animation team. Claude acts as the director: it writes a shot list, uses an image model (gpt-image-2) to build each still frame, uses a video model (Seedance 2.0) to animate each frame, then stitches the clips into one moving scene. A full multi-scene story drafts for about a dollar. Here is the exact workflow.

You can now make full motion graphics like this with Claude from a single sentence, for about a dollar a video. No animators, no expensive software. You have seen these animations everywhere lately: torn-paper worlds that move like stop-motion, clay characters at tiny clay desks, comic-book heroes with halftone dots and speed lines. All over Reels and TikTok right now.

The barrier used to be a paid tool and a skill set. It is not anymore. This is how the pipeline works and how to run it yourself.

What actually makes these animations

The key idea is that Claude does not draw anything. It is the prompt engine and the orchestrator. The visual work is done by two separate AI models that Claude chains together:

An image model builds a still frame in the style you asked for.
A video model takes that still and animates it into a short clip.

Then the clips get stitched into a continuous scene. That is the whole trick. Once you see it as “still, then animate, then stitch,” the mystery disappears and it becomes a repeatable format.

The creators pitching this usually use Higgsfield as the backend that holds both models. I wanted to own the whole thing, so I built a Claude skill that uses kie.ai instead. One API key on kie.ai unlocks both the image model and the video model, and Claude drives them.

The stack, named

Director: Claude, writing the shot list and every prompt.
Still frames: gpt-image-2, OpenAI’s image model, accessed through kie.ai. About five cents per frame.
Animation: Seedance 2.0, ByteDance’s image-to-video model, accessed through kie.ai. About a dollar per five-second clip at the high-quality tier.
Stitch: the clips are trimmed and hard-cut into one scene.

No After Effects. No render farm. No animation team. The most expensive part is a few dollars of compute.

The workflow, step by step

Here is the loop, repeated once per shot.

1. Turn an idea into a shot list

You give Claude a plain-English idea and a style. Claude breaks it into beats. For a short narrative, four beats is the sweet spot. A “how rockets work” explainer becomes: fuel, ignition, liftoff, space. A goal celebration becomes: the strike, the crowd, the knee-slide, the trophy.

2. Build the still for each beat

Each shot gets a still prompt written in the chosen style. The single most important line to include is this: “keep each object separate and clearly defined, as if every piece could move frame-by-frame.” That instruction is what lets the video model animate individual pieces instead of warping the entire frame. Skip it and your paper rocket melts instead of launching.

gpt-image-2 returns a public image URL. That URL feeds straight into the video model, no download or re-upload needed.

3. Animate each still

Each shot also gets a motion prompt. This describes only what moves plus the cadence, and it suppresses sound and text so the clips stitch cleanly. A good motion prompt ends with “no music, no voiceover, no on-screen text changes.” For stop-motion styles you also name the cadence, for example “around 8 to 12 frames per second, jumpy and imperfect, not smooth digital motion.” That is what gives it the handmade, stepped feel instead of glossy AI video.

Seedance 2.0 turns the still into a five-second clip. One note that will save you a failed run: the model rejects a three-second duration. Render at five seconds, then trim in the edit.

4. Stitch the beats into a scene

The four clips get trimmed to about two and a half seconds each and hard-cut together into one continuous roughly ten-second scene. Keep the same subject, kit, and palette across the beats and it reads as one story, not four disconnected loops.

The draft-then-final trick that keeps it cheap

There are two tiers of the video model. The draft tier costs about nine cents per clip and moves loosely. The final tier costs about a dollar per clip and follows direction far better. The workflow is simple: draft the whole scene cheap to lock the composition and motion, look at it, then re-render only the shots you are keeping at the final tier. A four-beat scene drafts for about a dollar and finals for about eight.

One skill, seven looks

The engine does not care about style. The style lives entirely in the prompts. The same pipeline produces:

Paper cutout collage: kraft paper and torn edges. Best for diagram-style explainers.
Claymation: plasticine with visible thumbprints. Characterful, great for stories.
Spiderverse anime: halftone dots, chromatic aberration, comic speed lines. For hype.
Isometric diorama: a tilt-shift toy world. The most reliable style and the best for scenes.
Needle-felt, liquid chrome, inflatable balloon: cozy, premium, and playful.

Same prompt structure, completely different world. To add a new style you write a new prompt recipe. The engine stays the same.

Where it shines: explainers

The format that works best is educational. “How does X work” gives each beat a reason to change. I built a paper-diagram “how rockets work” sequence that goes from a fuel cutaway to ignition to liftoff to space, with little cut-paper labels and action-reaction arrows. I built a tilt-shift “how gravity was discovered” scene with Newton under the apple tree, the falling apple, the lightbulb moment, and a tiny brass model of the Earth and Moon in orbit. Each one is a few clips stitched into a story for about a dollar to draft.

This is the same instinct a small brand can use. If you run an e-commerce shop like my jewelry brand Mogano, you can turn “how our rings are made” into a paper-craft explainer for the cost of a coffee, instead of booking a studio.

Keep it clean

Two quick guardrails. First, if you want several clips to look genuinely different as overlays, change the world, not just the style. The same scene in seven styles still reads as one thing. Different subjects read as distinct. Second, keep people faceless or generic and keep any trophies or logos generic, so you stay clear of likeness and trademark issues.

Frequently asked questions

Do I need to know how to code? No. Claude writes the shot list and every prompt from your plain-English idea. You describe the scene and the style.

How much does one video cost? About a dollar to draft a full multi-scene story, and around eight dollars to render the final version at the highest-quality tier. A single still is about five cents.

How long does it take? Each clip renders in roughly three to four minutes. A four-beat scene runs several clips in parallel, so a draft is ready in a few minutes.

Is this the same as Higgsfield? Same idea, different backend. Higgsfield is what most creators use. This is a self-built version on kie.ai, which gives you your own control over the models and the pipeline.

What styles can it do? Paper cutout, claymation, Spiderverse anime, isometric diorama, needle-felt, liquid chrome, and inflatable balloon, all from the same engine.

The takeaway

Stop-motion motion graphics stopped being a studio job. The whole thing is now: describe the scene, let Claude direct two models, stitch the result. The barrier is not the tool anymore, it is knowing the two prompt rules and the draft-then-final rhythm. Now you know both.

If you want the exact skill and setup I use, that is the deliverable. Grab it and build your first scene today.