Contents
0%Hey!
Claymation ads work stupidly well right now, and the whole thing comes down to what order you build them in. Here is one, then every step behind it.
Watch the ad
That is the opening of a keto recipe ad for a Thermomix audience: the hook, the guilt beat, and the agitation. Every frame is generated, every clip is animated, and one character carries the whole thing. The voiceover is Polish because that is the market it runs in, but the workflow is identical in any language.
The order matters more than the tools
Most people start by generating pretty clay images and then try to bolt a voiceover on afterwards. That is backwards, and it costs you hours in the edit.
The order that works: script, voiceover, storyboard, character, scene frames, video. Everything downstream of the voiceover gets timed against it, which is the entire trick.
Generate the ENTIRE voiceover before you touch a single visual, then have your AI analyze it. Voiceover pacing is never linear, some lines rush and some drag. If your AI assumes it is linear, every clip length it plans will be wrong, and you will spend your evening slowing clips down in your editor and regenerating shots.
Step 1: Write the script from real customer research
I am not going to pretend the style is what sells. A claymation ad with a weak script is just a nice looking ad that does not convert.
Before anything else, research the customer on Reddit, in reviews, in forums, in Facebook groups, in your own TikTok comments. The goal is to understand your buyer better than they understand themselves, so you can name the failed solutions they already tried.
Then break the script into named parts. Mine were hook, agitation, turn, proof plus vision, CTA. That structure is what the storyboard hangs off later.
Step 2: Generate the voiceover, then cut the dead air
The same voiceover before and after removing dead air: 1:08 down to 0:54
AI voiceovers come out with a lot of silence in them. This one had 13.8 seconds of dead air in a single take, and cutting it took the read from 1:08 to 0:54.
You would strip that in your editor anyway, so strip it first. Now the AI is analyzing the real final pace of the ad rather than a padded version of it, and every clip length it plans afterwards actually matches the cut.
Step 3: Pick your claymation style
The four claymation sub-styles side by side: cinematic, goofy, clean and simple, stop-motion puppet
There is not one claymation look, there are four that hold up, and picking the wrong one for the brand is a real mistake. The skill shows you this grid in the chat and recommends one for your product.
- Cinematic is the default for most ads. Macro lens, real fabric clothing, strong bokeh, premium feel.
- Goofy is oversized heads and googly eyes. Comedy beats and high energy hooks.
- Clean and simple is smooth matte forms and dot eyes. Minimal brands, calm explainers.
- Stop-motion puppet mixes real craft materials into the clay. Story driven, moodier, premium narrative.
One style per ad, locked for every prompt in that ad. Mixing them mid-ad is what makes generated video look cheap.
All four style blocks are packaged as one SKILL.md. Drop it into Claude, or run it inside Starpop, and it writes the prompts for you. Download the free claymation skill file.
Step 4: Lock the character before you generate anything
The hero character reference sheet, restated in full in every single prompt
Nine times out of ten your ad has at least one recurring character. Mine was Kasia.
The character sheet above gets restated in full, in every prompt, and the locked image gets attached to every generation after it. Abbreviating that description to keep a prompt readable is the fastest way to get a different face in shot four.
Whether you are on Starpop, or driving a generation MCP from Claude, you have to explicitly tell it that the character reference must be attached to every image AND every video generation. It will not assume it. Skip this and you burn credits on a character who changes face between beats.
Step 5: Storyboard every beat against the audio
The 15 part storyboard: each part with its timestamp, voiceover line, visual concept and duration
Because the voiceover already exists, the storyboard can carry a real timestamp and a real duration on every row. This ad came out as 15 video parts.
This is the stage that decides whether the ad is good. If you only pay attention to one thing here, make it the hook.
The hook is worth the extra hour
The script line was that sticking to keto has nothing to do with willpower. So the question became: how do you actually show willpower?
Four generated variants of the hook frame: a tug of war with an anthropomorphic cake monster
A tug of war against an anthropomorphic cake monster, staged inside a thought bubble above her head. Four variants, then one small edit pass on the winner rather than a full regeneration.
When a frame is 90% right, use the edit tool on it. Rerolling the whole prompt throws away the parts that already worked, and on a character shot you are gambling the face every single time.
Step 6: Generate the scene frames
Once the storyboard is solid, this part is boring, which is the point. The chat writes every prompt and you review rather than author.
Scene frame: the gathering, friends eating around a kitchen island
Scene frame: unboxing the Thermomix in warm golden lightScene frame: the gathering, friends eating around a kitchen islandScene frame: unboxing the Thermomix in warm golden lightIf the storyboard was done properly, the image stage flies. Every miss at this stage traces back to a vague beat description, not to the image model.
Step 7: Animate, and let the model do the transitions
There are two ways people do this, and they are genuinely different workflows.
The start and end frame chain. You give a first shot a start frame and an end frame, then reuse that end frame as the next shot's start frame, chaining the whole ad together. Kling 3.0 is the usual pick for it. It works, and it is a lot of frames to manage.
Multiple reference images in one generation. This is what I did. You hand the model several scene frames at once and describe how it should move between them, and it resolves the transitions internally.
The real video generation: the motion guard prompt, five reference images, and a 12 second output
That is the actual prompt panel. Five reference images, each explicitly labelled with which beat it is, a 12 second duration that came from the audio analysis, and a motion guard at the top that keeps the clay looking like clay:
Handmade stop-motion claymation throughout, matte plasticine, fine tool striations and thumb-worked surfaces, slightly stuttery 12fps stop-motion cadence, no motion blur. Non CGI. Non cartoon. Animation must start at the first frame.
That is the raw 12 second output, four beats, generated once. No rerolls.
The generation record: Seedance 2 Mini, five inputs, 12 seconds, 720p
Good images plus a specific transition plan is the difference between one generation and six. Premium video models are expensive enough that rerolling is a real cost, so spend the effort on the storyboard and the frames instead. Every clip in this ad worked on the first try.
Step 8: Assemble against the voiceover
The last part is the easy part, precisely because of step 2. The clips were planned against the dead-air-removed voiceover, so they drop onto the timeline and line up.
The finished timeline: generated clips stacked above the dead-air-removed voiceover track
That is the whole edit. The clips sit above the voiceover track, which is still the dead-air-removed file from step 2, with a music bed underneath.
No slowing clips to fit. No regenerating a shot because it came back three seconds short.
Next steps
The whole thing is one skill file plus one chat. Build in that order: voiceover, storyboard, character, frames, video.
The full production order, the tactile style lock, the prompt structure, the continuity rules and the failure fixes, in one public SKILL.md with no account required. Get the free claymation skill for Claude.
If you want the prompt level detail, the four locked style blocks and the motion guard are broken down in the AI claymation prompt guide. And if you want to see what the format looks like across different brands first, there are five claymation ads worth studying here.
