
I am making my own game, Noční linka (Night Line), and Claude Opus 5.5 wrote the 3D models of its world as scripts for Blender. GPT-6 Astra then reviewed them independently and fixed animations. This article rests on what I saw along the way: what the AI did, what failed, which gates and criteria I set, and what the result looks like. An overview of what is happening around AI and games (Genie, Steam, developers) comes at the end, verified on 20 October 2026.
Can you make a game with AI? In my case yes, as an unfinished prototype: AI wrote scripts for 3D models and fixed animations, but what counted as good was decided by an independent check of real renders and by a human. Some of the fixes failed (the character's coat ended as HOLD after three rounds), and that is part of the process. The game did not appear from a single brief but in small steps with review. Generated worlds (Google Genie 3, World Labs Marble) are still demos and 3D scenes, not game engines. On Steam you do not have to disclose tools that only speed up development, but you must disclose AI content in the game and its materials, or content generated at runtime.
- The world of the game from scripts. The panel blocks, village houses, train, cars, bus, street furniture, mill and gamekeeper's lodge were written by Claude Opus 5.5 as Python scripts for Blender (commits from 7 and 9 October 2026). Blender 4.1.1 ran windowless from the command line, rendering with Cycles on the CPU, with no Blender MCP and no text-to-3D generator.
- An independent audit. GPT-6 Astra (10 to 11 October 2026) praised the bus's readable silhouette, frames, handrails and seats, and criticised bare sides, flat doors and roofs that are too regular on the village houses.
- Animation fixes. The mop and the pole grip on the bus passed in two rounds (palm anchor deviation of at most 4.54 mm and 2.19 mm, measured on bones, not on the full mesh). The coat for a female character stayed HOLD after three rounds and was not enabled in the game.
- What I do not have. A direct comparison of Opus and Astra on the same task did not happen, and I did not measure how much of the work the AI did in the game.
What is Noční linka and how do I make it with AI?
Noční linka (Night Line) is my own unfinished game, not a client job. You drive the night bus on line 913 through the fictional Bořín district in 1990s Czechia: stops, passengers, haggling at the door and troublemakers. It runs on Godot 4.7.1 and third-party assets are CC0 only. It is an alpha and a prototype, and the game has not been released.

Who does what and by which rules
The author is an AI agent, either Opus 5.5 or GPT-6 Astra, and an independent critic works beside it, judging the real renders, not the author's claims. Several agents work on the game in parallel, each in its own working copy of the repository, and the heavy processes (Blender without a window, Godot without a window) run one at a time. They hand finished work over in a note with the exact scope, what passed and what did not. I describe the general principle in the article on AI agents and the harness.
Every asset passes eight production gates:
- reference and purpose,
- a spatial scale check,
- model,
- materials,
- skeleton and animation,
- GLB export with a reload test in Godot,
- a gameplay check,
- independent critique.
For every finding the critic must give the screenshot file or the sequence time, the reason and a condition for acceptance. Where it lacks a sufficient view or motion, it writes NEOVĚŘENO (not verified). The acceptance criterion should then not be "looks good" but, for example, a grip deviation in millimetres for three character heights.
3D models from scripts
The panel blocks, village houses, train, cars, bus, street furniture, mill and forester's lodge were written as Python scripts for Blender (bpy) by Claude Opus 5.5. Blender 4.1.1 ran headless, rendering with Cycles on the CPU. The advantage: the model's source is a readable script, so I can fix one detail without regenerating the whole scene. These are prototype-level game models. The comparison of scripts with MCP and generators is in Blender and AI.

What did the AI do in the game and how did it turn out?
A summary of my work with two models, details are in the article GPT-6 Astra:
| Job | Who | Result | Note |
|---|---|---|---|
World models from scripts |
Opus 5.5 |
prototype-level models |
the audit criticised bare house sides, repeated stains and plaster outlines |
Audit of the models from renders (10 to 11 Oct) |
Astra |
findings, fixed nothing |
it did not recommend replacing the models wholesale |
Mop |
Astra |
accepted, 2 rounds |
palm anchor deviation at most 4.54 mm across 73 frames |
Punches (jab, cross, hooks) |
Astra |
accepted, 2 rounds |
sideways wrist travel at most 0.252 m instead of 0.376 m |
Holding the bus pole |
Astra |
accepted, 2 rounds |
heights 1.60 / 1.75 / 1.90 m, deviation at most 2.19 mm |
Coat for a female character |
Astra |
HOLD after 3 rounds |
gap at the waist, elbows clipping, not enabled in the game |
The animation fixes ran on 11 October 2026 (batch V3, Blender 4.1.1 without a window, at most three preview rounds per item, and all 35 of 35 previews were actually opened and inspected). On the bus the audit praised the readable silhouette, frames, handrails and seats. On the village houses it criticised bare sides and rear faces, flat doors and overly regular roofs: up close they look like a kit. On the panel blocks, entrances, loggias and antennas work, but the dirt map repeats.
An important limit: contacts were measured on the character's bones (a proxy), not with a full mesh intersection test. "Accepted" means an accepted source and previews, not finished gameplay. The animations are not yet wired into the game and the in-game check of passengers in Godot remains. A screenshot also proves the look of one pose, not the motion.
What did not work?
The coat for the female character got real sleeves, a longer hem, a collar and buttons, but after three rounds a gap at the waist remained and the elbows passed through the body. An independent critic confirmed the HOLD verdict and the coat was not enabled in the game. Better a HOLD than an accepted defect.

The follow-up batch of animations, V4 (nausea, swaying, provocation, calming hands), ended up all HOLD after the first round: a hand was at the temple instead of in front of the mouth, an index finger stayed curled and the fingers were too clenched. A record of the next phase from the same evening says that after further rounds the inspected previews passed and the assets loaded in headless Godot, but wiring into the game and contact on the characters remain HOLD. A direct comparison of Astra and Opus on the same task did not take place, so I do not claim that either is better.
What does the result look like so far?
Before and after: the bus headlights
Two screenshots from the bus cab on a night street: on the left before the headlight adjustment, on the right after it. The clock in the cab reads 23:03 on the left and 23:06 on the right, so it is not exactly the same moment. The lighting of the street and of the figures in the shot differs between the two. What exactly changed in the adjustment I do not describe here, because I cannot read it from the image and I do not have it documented.

I use the same method as for the models: same place, same angle, a screenshot before and after, and a verdict only from the image. According to my workflow, light is tuned in the engine, not in Blender, because Blender may render differently.
How do you start your own game with AI?
- Small scope. One mechanic, one scene. Without tests and review, you get the mistakes from Vibe coding: 7 amateur mistakes.
- An engine where you can verify the result. I use Godot, where a model can be reloaded and checked even without a window.
- References before models. The OpenAI case study starts with requirements for the player experience and image references.
- A measurable acceptance criterion. For the pole on the bus it was grip deviation in millimetres for three character heights.
- An independent critic and a round limit. The author does not judge itself, the critic gets the real renders, and after three rounds a human decides.
- Asset licences. Third parties CC0 only. For generators, read the terms: outputs of the free Meshy plan are under CC BY 4.0 (verified 20 October 2026) and the Hunyuan3D 2.1 licence expressly does not apply in the EU.
- A record for Steam. Keep a list of what AI produced.
The general workflow for AI development is in How to build an app with AI.
What is happening around AI and games? A short overview
This part is informational, not my experience. The data comes from the linked sources, verified on 20 October 2026.
Worlds from a prompt. Project Genie is an experimental app built on the Genie 3 model. Google launched it on 29 January 2026 for Google AI Ultra subscribers in the US aged 18 and over, generation is limited to 60 seconds, and Google plans to expand to more territories. According to Google, the worlds may not look realistic, may not follow the prompt or physics closely, and character control can have higher latency. Marble 1.1 from World Labs (2 April 2026) creates 3D worlds you export as splats or meshes. That is a basis for a scene, not a game engine: goals, rules and progression must come from somewhere else.
Steam. Valve changed its AI form on 16 January 2026, summarised by Game Developer:
| Situation | Disclose on Steam? |
|---|---|
| A code assistant or another tool that only speeds up development | No |
| AI content in the game and in related materials, such as the store page | Yes, described in a text box |
| The game itself generates AI content at runtime | Yes, by ticking a box |
Where exactly the line lies for models written as scripts, each publisher has to judge from the current form. Noční linka has not been released. A list of what the AI produced is worth keeping as you go; for my game it would include, for example, the 3D models written as Blender scripts.
The rules change and this is not legal advice. What counts is the wording of the form in Steamworks on the day you submit the game.
How developers feel. According to the GDC 2026 survey as summarised by the press (over 2,300 professionals), 52% of developers view generative AI negatively (18% two years ago), 7% view it positively (13% last year) and 36% use it at work. You have to expect that some people will react to AI in a game with suspicion, which is why I write openly about using it.
I build Night Line for myself, not for a client. For clients I make short educational games for schools and publishers; demos such as Wiring a socket or Gear ratio are on the Educational games page. If you need an app or automation where an agent works and independent review verifies the result, see Apps and automation made to measure.
FAQ
Can you make a game with AI?
Yes, as a prototype, if you work in small steps. An AI agent writes code and some of the assets (for me, 3D models as Blender scripts), but a human does the brief, the selection of results and the check of playability. My Noční linka is still an alpha and has not been released.
How much of the work does AI really do in a game?
I have not measured the share and do not want to guess percentages. In my game AI wrote scripts for 3D models and fixed animations, but an independent critic judged the results and some fixes failed: the character's coat ended as HOLD after three rounds. Review is part of the work, not an add-on.
Do I have to disclose AI on Steam?
Under the form change of 16 January 2026, you do not have to disclose tools that only improve development efficiency, such as a code assistant. You must disclose AI content in the game and its materials, and content generated at runtime. Check the current form in Steamworks before submitting.
What is Google Genie 3 and can I use it in the Czech Republic?
Genie 3 is a world model that powers the experimental Project Genie app. As of 20 October 2026 Google offers it to Google AI Ultra subscribers in the US aged 18 and over, a session lasts at most 60 seconds, and Google plans to expand to more territories. It names only the US among the supported countries, so check availability in the Czech Republic before you try.
Which engine should I choose for a game made with AI?
One where you can verify the result. I use Godot 4.7.1, where a model can be reloaded and checked even without a window. For Unity, the editor assistant Aura for Unity came out on 14 May 2026; I do not assess it, as I build in Godot. OpenAI's study builds a browser game in TypeScript with Three.js. What matters is whether the agent can run the result and compare it with the brief.
