AI 3D models let a solo programmer design, rig, and program a complete web fighting game in a single week. The build worked; the audience test did not. Eleven AI-generated fighters became an auto-battler roguelike that held test players for roughly one minute on Poki. This article walks through the pipeline, the shortcuts AI provided, the manual fixes it still required, and the data-driven conclusions.
What Is the One-Game-Per-Week Experiment?
The experiment is simple: ship one playable web game every week, then judge each one by platform data rather than personal taste. Brazilian developer and YouTuber André Elias, one of the roughly 600 developers publishing on Poki, the world's largest web games platform, set this rule after noticing that guessing what players want had never worked for him. Every game enters a test period where Poki's own analytics measure how long real players engage.
The first candidate was originally a virtual pet project called Don't Let Your Pet Die. Mid-sprint, a different idea took over: a fighting game starring well-known Brazilian developer YouTubers. That pivot produced the game actually shipped, an auto-battler roguelike, and it is the one whose results opened the experiment. The Brazilian dev-YouTube scene this game draws from includes channels such as Gustavo Dev Doido and his Bootcamp do Dev Doido community, plus course platforms like Crazystack TypeScript, which train exactly the kind of programmers who attempt projects like this one.
How Were the AI 3D Models Created?
The AI 3D models came from a three-step pipeline: an AI-generated reference image, mesh generation, then manual cleanup. The developer is a programmer, not a modeler, and hand-modeling in a traditional tool had always produced poor results for him. AI changed that equation in early 2026, when he revisited generation tools he had tested years before and found the output much improved.
He had already built two test models of YouTubers Mano Davi and Lucas Montano, posted the idea of making them fight on Twitter, and gotten an enthusiastic response. When the weekly experiment began, that idea became the game. In total he produced 11 fighter models, including himself: Mano Devin, Lucas Montano, Yud Ganeco, Felipe de Campos, Fiasco, Jovem Tranquilão, Fábia Quita, Diego Fernandes, Bero Dev, Augusto Galego, and André. He plans to add more characters in future updates, especially more women, which he admits the roster is missing.
The three stages, in order:
- Generate a T-pose concept image of the character with AI.
- Convert that image into a 3D mesh with an AI model generator.
- Import the mesh into Blender and fix the rigging and weight paint by hand.
Step 1: Concept Images With ChatGPT
Each character starts from a reference photo plus images of the desired result. The developer used ChatGPT, OpenAI's assistant, to generate a character image in a T-pose, the arms-out stance that makes skeleton placement far easier later. When his free image-generation quota ran out, he built his own generator: a simple Next.js app wrapping his own prompt, giving him unlimited generations without a per-image cap.
Step 2: Mesh Generation With Tripo3D
The finished concept image goes into Tripo3D, an AI 3D model generator the developer chose after comparing available options. He states plainly that this is not a sponsored placement, though he uses a referral link that grants him credits. The image-to-mesh step is where most of the traditional modeling time disappears; a usable character mesh comes back ready for rigging.
Step 3: Rigging and Weight-Paint Fixes in Blender
Bones define where each part of a model sits during an animation, and AI can generate them too. What it cannot yet do reliably is assign them correctly. In Blender, the open-source 3D suite, the developer spent significant time fixing weight paint, the mapping that decides which part of a mesh moves when a bone moves. In one model, the glasses deformed whenever the arm animated, which made no sense.
The honest takeaway from the whole pipeline: AI accelerates 3D character creation enormously, but human correction is still mandatory. A programmer who cannot sculpt can now produce game-ready characters, yet cannot skip Blender entirely. The same pattern repeated in code generation and UI design, where AI produced a usable first-pass interface but a worse final design than the developer's own rework.
Why Did a Fighting Game Become a Roguelike Auto-Battler?
The original plan was YouTuber-versus-YouTuber combat with ragdoll physics on knockout. The ragdoll work consumed so much time that it was cut. Meanwhile, comments on the previous video pushed a roguelike, a genre where each run differs from the last because players choose their own build, creating the classic just-one-more-run pull. That matched the experiment's goal of finding an engaging game model.
The result is an auto-battler roguelike: you assemble a team of fighters, buy items between rounds, and watch them fight on their own as difficulty climbs. The developer was playing HighGround-style titles at the time and cites Hman 2 as an inspiration. Combat mechanics included dodge chance, critical hits, stuns, and knockbacks. Roughly four pages of written planning covered every ability and item before implementation began.
How Did the Developer Keep AI-Written Code Performant?
A screen full of 3D characters can destroy performance, so the developer wrote performance and code-structure rules before letting AI touch the codebase. The rules document compiled every mistake from his previous game, Corte Kings, and was fed to the AI so it would not wreck the code or the frame rate. The result was a game that ran well and stayed maintainable.
The visual pass came last: improved lighting, a cartoon shader effect, a basic floor texture, hit effects, blood particles, music, and sound effects. That mattered because the plan was to keep shipping weekly. The lesson generalizes: AI coding tools follow explicit constraints far better than vague instructions, and a rules document is cheap insurance.
What Did the 500-Player Test on Poki Actually Show?
The game was submitted to Poki with a 500-player test cohort, and the developer hoped for around ten minutes of average engagement, since he had personally played it for thirty minutes at a stretch without noticing the time. The measured result was about one minute, roughly on par with his earlier games, including one built in only two days. More planning effort did not translate into more play time.
His theory is audience mismatch rather than a bad game. Poki's player base skews toward children and teenagers, who may not want a strategic, hands-off auto-battler where characters fight without direct control. He is skeptical of the claim that players simply did not recognize the YouTuber characters, arguing that a fun game survives unknown names. To test the hypothesis, he released the game to his own audience with analytics installed, tracking session length, character picks, item choices, and balance outliers. If the numbers improve, a Steam release is on the table, where revenue potential is higher but quality expectations rise with it, likely requiring more characters and items.
| Release target | Audience | Revenue potential | Quality bar |
|---|---|---|---|
| Poki test | Children and teens on the web platform | Low during test | Ship fast, simple games |
| Own audience link | Dev-YouTube viewers who know the characters | None directly | Must beat the one-minute benchmark |
| Steam | Buyers who pay up front | Considerably higher | More characters, more items, polish |
Weekly Poki releases continue in parallel in 2026, tuned toward simpler games that fit that audience, and a new game in that simpler style is already under way.
Frequently Asked Questions
What are AI 3D models?
AI 3D models are character or object meshes generated by machine-learning tools from an image or text prompt. In this project, Tripo3D converted concept images into rigged meshes, cutting out traditional modeling work almost entirely.
Can a programmer without modeling skills make a 3D game now?
Yes, with caveats. The pipeline of ChatGPT concept art, Tripo3D mesh generation, and Blender cleanup produced 11 game-ready characters in a week. Weight-paint corrections still demanded real Blender skill and time.
Is AI-generated game art good enough for release?
For meshes and skeletons, mostly yes after manual fixes; for game UI design, no. The developer replaced his AI's interface design with his own after finding it weak, especially for games.
Why did the game underperform on Poki?
The developer's own theory is audience mismatch: Poki's young players may not engage with a strategic auto-battler where fighters act on their own. He is testing this by releasing the game to his own audience with analytics and comparing the data.
What is an auto-battler roguelike?
It is a game where you build a team and buy items between rounds, then watch battles resolve automatically. Each run differs based on your choices, and difficulty rises until only a well-built team reaches the end.
Turn Your Own Dev Videos Into Written Articles
This experiment worked because a spoken devlog became something a reader could scan, quote, and search: the pipeline, the numbers, and the theory are now all on the page. If you have the same kind of knowledge sitting inside your own YouTube videos, a build walkthrough, an interview, a post-mortem like this one, you can turn it into a written article without writing from scratch.
Skala Blog does exactly that: paste a YouTube URL, get the transcription, and generate a structured article from it. The one-minute lesson from this game applies to content too, meeting your audience in the format they actually use costs far less than rebuilding everything from zero.
Fork this article
Start a new branch from the same video, shaped your way. You keep the credit; the original keeps the attribution.
A fork in another language is filed as a translation of this article, so the two pages point at each other. You can unlink it later from the editor.
0/240
You are creating
- Format
- For
- Language
- Source
- Your angle
You will be asked to sign in before it is generated.
Buy credits