Bonk & Zuzu: How to Make an AI Cartoon Series for Instagram
Bonk & Zuzu is a silent stone-age couple comedy made entirely in Kagelot. One month after the first post (3 October 2026), the Instagram account @bonk_and_zuzu had 74.7K followers and 29.5 million views in the last 30 days, and one clip, the dress, passed 8.5 million views.
Each episode starts as one plain-language idea in the chat. Kagelot writes the screenplay and the prompts, generates the images, attaches the right references and renders the video, with the same two characters in every clip.
Numbers checked: 3 November 2026 (followers, 30-day views) · 27 October 2026 (per-clip views).
How I Made a Viral AI Cartoon Series (Full Tutorial, No Prompting) · 8:22
Chapters
- 0:00A one-month-old account with nearly 30M views
- 0:26Copying a style from reference images
- 1:21Creating the two characters (and fixing them)
- 2:28Part 2 · The first clip
- 2:47The AI writes the whole prompt
- 3:10Where the files live: Files Explorer and canvas
- 3:43The cave image asset
- 4:17Generating the video
- 5:01Part 3 · The second clip
- 5:20AI memory: it reuses the cave on its own
- 5:44Why the drawings are separate images
- 6:40Saving the session to project memory
- 6:58Part 4 · It learns what works
- 7:14Telling the AI what went viral
- 7:44Stuck on ideas? Ask
Step by step
1. Copy a visual style from a few reference images
Attach a few reference images in the chat and ask Kagelot to copy the style. It describes the look in words and saves it in the Style Library ("Uncanny Caricature 3D"), so every image in the project follows it.
2. Design the two characters and lock them
Pick Viral mode to shape the series idea, then ask for each character. Image Gen keeps every version, newest on top. When one is right, use Add to character: the approved sheet (REF-bonk-sheet.png, REF-zuzu-sheet.png) is now attached to every image and video prompt that names that character.
3. Write the idea in one line; the AI writes the prompt
Switch to Direct and describe the clip the way you would tell a friend. The AI already knows the story and both characters, so it writes the screenplay, the three phone-camera positions and the full Seedance prompt (about 3,300 characters for Clip 1). You never describe clothes, materials or the location.
SHOT 1 — HANDHELD MEDIUM WIDE FROM ACROSS THE CAVE, PHONE AT CHEST HEIGHT WITH SLIGHT NATURAL SWAY: The cave from [@image1]. [@image3] sits on the low flat boulder, her back turned to [@image2]. [@image2] stands behind her and scratches her back with his tiny hands. She points over her shoulder to one spot. He scratches a completely different spot.
4. Check the files: Files Explorer and canvas are the same project
Everything the AI writes is saved as files. In the Files Explorer the clip has its own folder with its prompt file; on the canvas the same folder is a clip bar with its images and videos. Open the prompt file to read or edit it, or just close it and generate.
5. Generate the image assets, with no image tagging
Click the clip bar, then Generate → Generate image assets. The image card appears in the chat with its prompt already written: pick the model (GPT Image 2.5) and one variation, generate, open it full size in Image Gen and assign it to the clip. The prompt file already says which image goes where, so you never tag references.
@image1 → CLIP-02-asset-cave-interior.png @image2 → REF-bonk-sheet.png @image3 → REF-zuzu-sheet.png
6. Render the video
Click the clip bar again → Generate video ("All 3 references ready"). The video card appears in the chat: pick Seedance 2.5 and Render. The finished clip lands in the clip's videos folder.
7. Reuse locations with project memory
For the second clip you never mention the cave. The AI remembers where the characters live and puts the same cave image into the new clip as @image1. New things (his drawings) become separate image assets, so the cave stays identical and only the drawings change.
@image1 → CLIP-02-asset-cave-interior.png @image2 → REF-bonk-sheet.png @image3 → REF-zuzu-sheet.png @image4 → CLIP-03-asset-art-single.png @image5 → CLIP-03-asset-art-wall.png
8. Save the session and teach it what works
Click the save icon in the chat: the session is written into memory.md, so the next session knows the style, the characters and the locked locations. When a clip goes viral, tell the AI. It logs the numbers in memory, works out why the clip worked and builds the next ideas on that pattern.
What went wrong and how it was fixed
- Bonk's first design came out too realistic. The cause was the style file itself; it was fixed and Bonk regenerated. The wrong versions stay below the new ones in Image Gen.
- The round, pot-belly Bonk was replaced by a lean design: stone-age people were not fat.
- The first take of "The Portrait" had the wrong camera. One message re-directed it: the camera behind him while he paints, then a tilt up to the wall of rejected drawings.
- A batch of new ideas "didn't feel the same as the dress clip". The AI went back to the pattern from that clip (a real couple cliché, cut to black, the next-morning reveal) and rebuilt the ideas on it.
FAQ
Do I need to write prompts?
No. You describe the idea in plain words; the AI writes the screenplay, the camera and the full prompt. You can open the prompt file to read or edit it.
How do the characters stay the same in every clip?
Each character has one approved character sheet. Every prompt refers to the character by name and Kagelot attaches the sheet automatically.
Do I have to tag reference images?
No. The prompt file holds the mapping (@image1 is the cave, @image2 is Bonk, and so on) and Kagelot attaches the right images when you generate.
Which models does this tutorial use?
GPT Image 2.5 for images, Seedance 2.5 for video, and Claude Sonnet as the chat model.
Does the AI remember the project between sessions?
Yes. The project memory (memory.md) keeps the style, the characters, locked locations and which clips performed well, so you never re-explain them.
Transcript
Read the full narration
Part 1
This Instagram account is one month old. Almost seventy-five thousand followers, and nearly thirty million views. I didn't film anything, and I didn't draw anything. Every video came from one tool, Kagelot, and a conversation with an AI. The biggest one got eight and a half million views on its own. This is how I made it, from the very first step. First I need a look. I found a few pictures of a 3D caricature style I like. I upload them, and I write: copy this style. It studies the pictures and writes the style down in words. It's saved in my Style Library, and every image in this project follows it. Next to the chat box there's a mode menu. Viral is for ideas, Direct writes a clip shot by shot, Screenplay writes the script, Design makes the images. Then the idea. I pick Viral and describe the show: a husband and wife from the stone age, doing daily things, funny and silly, no talking, like Mr Bean. Fifteen to thirty seconds. It saves a series document in my project, with the format, the rules and a list of episode ideas. Now the husband. I ask for a catchy name. It picks Bonk: one syllable, the sound his head makes when something lands on it. The first pictures are close, but not right. So I tell it, the way I'd tell a person. It finds the real cause. The style file said: realistic skin, visible pores. It rewrites that line, and makes the hands and feet tiny. Two more notes: cover his chest, and make him look thirty, not old. Then I launch the card. Then the wife. It names her Zuzu, and makes her tall and thin, the opposite of Bonk, so they never look alike. One thing still bothered me. People in the stone age weren't fat. So I ask for a second Bonk, and I pick the lean one. I pick my two characters, and now they're locked. Every image and every video after this uses them. In the next part, we make the first episode.
Part 2
Now we're going to make our first clip. I switch the mode to Direct, and I pick Claude Sonnet as the AI. I just write my idea the way I'd say it. And I tell it we need an image of their home, a cave. This is the main thing to understand. The AI already knows the story and both characters. So I don't write the prompt. It writes the screenplay, the camera, and the full prompt by itself. I never say what they wear, or what the clothes are made of. It already knows Bonk from his character sheet: the hair, the fur, the face. Everything it writes is saved in the project. In the Files Explorer, the clip has its own folder, with the prompt file inside. On the canvas, the same folder shows up as a clip bar, with its images and videos. Same files, just the bigger picture. I open the prompt file. This is the whole prompt the AI wrote, and even which image goes where. I can edit it, or just read it. If it's good, I close it and click Generate. No image tagging, no prompting on my side. First we need their home. I click the clip, then Generate, then Generate image assets. The image card shows up in the chat, with the prompt already written. I pick the model, GPT Image, and one variation. Generate. In Image Gen I open it full size. The rock walls, the fire pit, the stone seat. I didn't describe any of it, and it's right. Then I assign it to the clip. Now the cave is part of clip one. Back to the clip. Generate, Generate video. All three references are ready: the cave, Bonk and Zuzu. Kagelot already knows which image goes where. Zero tagging. The video card shows up in the chat. I pick Seedance 2.5 and click Render. Same faces, same hair, same clothes as the character sheets, in their cave. That's our first clip.
Part 3
Part three. The second clip, and this one is more complex. I pick Direct and write my idea, the same way as before. The clip bar lands, with the screenplay and the prompt ready. Now look at the memory. I never mentioned the cave. The AI remembers they live in the cave from clip one, so it puts that same cave image straight into this clip. I don't search for it, and I don't generate it again. The cave is locked for every episode. This time it asks for two new images: his first drawing, and a wall full of the drawings she hated. Why separate images? Because the cave has to stay exactly the same. The cave comes from the old image, and only the drawings are new. The AI understood that, and wrote the prompt to fit. Generate image assets. Two cards, one for each drawing. Generate. They land in Image Gen, and I assign them to the clip. Generate video. Five references, all ready. Render. The same cave from clip one, the same two characters, and his drawings on the wall. One last thing. I click the save icon, and the AI writes this session into the project memory. So when I close the project and come back tomorrow, it remembers where we left off. I never have to explain who Bonk and Zuzu are again.
Part 4
Part four. This is every clip so far, all made the same way, in one project. And this is the part most people miss. When a clip goes viral, I tell the AI. It logs the numbers in the project memory, and works out why that clip worked. The dress clip was the first one that went super viral. It's past eight million views now. Every hit teaches it what my audience likes, and the pattern stays in memory. So when I'm stuck, I just ask for ideas. If the ideas miss, I say so. It goes back to the pattern from the dress clip, and builds new ideas on it. And when I'm stuck on the shots, it helps there too. I bring the taste. It remembers everything. That's how I run a whole show with Kagelot.