Shut. The. Hell. Up.https://t.co/QVEzkIvXGv pic.twitter.com/mr5vIeJmf2

— Mikkel Malmberg (@mikker) August 12, 2026

This tweet got popular. It even brought out a hater or two demanding the receipts. So, without too much talk and definitely with no intention of selling you a course – here's the entire saga and agent transcripts that got me there.

I won't call it a workflow because it's all just made up as I go. There are tools and techniques that stay but mostly, you just talk to the models.

It all started when I was talking names for my Rails UI kit with my internet friend Marc.

This was november 2024.

Prompting in MidJourney was and is a very explorative process. It's about prompting something then prompting the prompt. Using the output as reference images for the next rounds of prompts. Until something badass randomly comes out.

The very first prompt was this and an RC car I found through Google Images:

minimal icon of a remote controlled buggy, vector style, yellow car, hot red background

I then tried the Niji 6 model instead and the slightly updated:

minimal icon of a remote controlled buggy with a giant jet engine on top, vector style, yellow car, hot red background

Something about that one output.

The red circle background was added manually by me. That's where the car originally came from.

Fast forward two years. And. Yeah, the forward has been very fast these two years.

I look at my UI kit again. AI and agentic coding has changed everything.

I could use an updated illustration for the updated 2.0 version.

That's the original car image + a character sheet of a stick figure guy I had made for a talk and used again for an illustration.

Alright, so I'm getting near something I quite like here. Not perfect but the composition is very cool.

I have an idea for another angle that would look awesome as the website footer maybe.

Huh. Clearly the model (gpt-image-2) is confused about the perspective here. I wonder how I could help it, I think.

ChatGPT can talk for very long without loosing its mind but image generation usually only gets worse and worse the longer the thread becomes (in my experience.)

In a new thread, I try the character sheet method that had worked for me before.

Take it back to the other one...

Better but still not it. The concept of a jet engine on the roof is clearly not something it has considered before.

I keep thinking: I know this is what I want. I'm convinced it'll look cool. How can I help the model get it right?

I remember having seen tweets about connecting coding agents to Blender via MCP. Surely, if I can model the car in actual 3d space, we can get a picture – or at least a draft composition – of it from any angle.

I open up codex. Ask it how do I connect you to Blender via mcp? Few different ways, it says. One is literally just uvx blender-mcp. Perfect for the amount of effort I'm willing to throw at this. Here's the actual thread transcript

Open Blender.app. Run uvx blender-mcp. Enable shit. They find each other.

Give Codex the character sheet and tell it to model it in Blender. Here's the actual thread.

Use blender mcp to model this car [Image #1] Make it looks as close to the reference image as possible. Feel free to imagine parts not visible on ref sheet

"More detail"

I wasn't too impressed with Codex' results. Or. I was impressed by any of this working at all! But the results didn't look like the car I knew and loved.

At least we could now generate a sketch of the exact angle we wanted.

Not bad now!

I wonder though... What if we gave the task to Fable?

Oh, a few weeks before this, I had already tried making ChatGPT create a fake photo shoot of the car. I gave it that as well.

Here's the very first, single prompt result:

Insane. Here's the whole thread. It goes all the way to the gamified version.

Try expanding the tool calls there. You can follow the whole process as the agent contemplates its work from different angles. So interesting!

After a few rounds of back and forth, we arrive at the dusk scene from the original tweet.

Already here, this is clearly just completely bonkers that a statistics based text generation machine can build this.

Through the medium of language (ish), it has to understand me. Understand the images I give it. Understand the concept of 3d space and the concept of apps and the concept of modeling the former in the latter.

Just. Completely. Bonkers.

Of course, we aren't even done.

Can you build it in Three.js I ask it. It can.

Make it animated, I say. Make the engine switch on and off.

It's pretty good!

I see a tweet of someone using Blender to model and choreograph a scene then Seedance 2.5 to make it realistic. I wonder ...

First output isn't very good. I passed it a recording of the Three.js scene as reference. Maybe I did it wrong?

Basically just a copy of the input.

How can I help it? I think. Maybe if I generate a still of a realistic render, tell it to use that, AND a video of the low-poly choreography, tell it to use that too.

So, off to ChatGPT again.

Exactly what we want!

Back to Seedance and Replicate.

And there we go. The final, tweeted, jawdropping version.

The Claude thread shows the rest of the path to the playable version which was basically asking it to make it drivable and that's it.

I hope this post showed you that …

1. the future is here and it is bonkers
2. there's no magic. It's just a matter of thinking what's my next step and playing yes, and with whatever output you get.

Have fun!