How to make a game like GTA with AI

There is a playable answer to this question on OrigoZero. Grand Theft Auto VI, PS1 Demake is a compact open-world crime game: a Vice City district you can drive around, a liquor-store score, police chases, a stash-house raid and a 90-second reveal trailer rebuilt shot by shot, all rendered the way a 1998 PlayStation would have drawn it. It runs in your browser. An AI agent built it in two days, and it left its working notes inside the world.

This post is the story of that build, taken from the agent's own commit history and design documents. The prompts are quoted as they were recorded. The parts that are still rough are listed too, because the agent listed them.

What "a game like GTA" means here

The genre is a third-person open world where you walk, steal a car, commit a crime, get chased by police and either lose them or wake up at the hospital. That is the whole brief. Everything else in the world (the two protagonists, the trailer, the pink and teal palette) is fan interpretation of a game that is not out yet, and the world says so on its own page: an unofficial fan recreation with generated art and simplified systems.

The demake angle was a deliberate choice, and a smart one. A PS1 look sets the bar where an agent can clear it in a day: coarse textures, angular models, a 320 by 240 raster and flat baked light. Nobody expects a fan demake to look like a hundred-million-dollar production, so the result reads as intentional rather than unfinished.

The first prompt, and the first correction

The build started with a scope prompt. The agent recorded the correction it received almost immediately in its spec file:

The user corrected scope: actual GTA VI devolved, not generic PSX open world. Lucia Caminos and Jason Duval, modern Leonida / Vice City.

That correction did a lot of work. "Generic retro open world" is what an agent produces when the brief is vague. Naming the two characters, the city and the era pinned every later decision to something recognisable.

The art direction was one paragraph, and it is the single most reused prompt in the whole build:

GTA VI demade as an authentic 1998 PlayStation 1 3D game: coral pink, sunset orange, faded teal and deep violet palette, coarse 64px painted textures on angular low-polygon models, flat baked dusk light, ordered dithering, jagged 320x240 edges, sun-bleached stucco and scuffed car paint.

Four reference images were generated from that paragraph first: an in-play frame that set the camera and street art, a title composition, the two protagonists' silhouettes, and the red getaway coupe. The agent then read those images back while building, so the 3D scene had a fixed target instead of a fresh interpretation every time.

How the agent organised itself

The spec file shows the agent splitting the world into three owned areas and handing each to a sub-agent with explicit boundaries:

• City owned the streets, buildings, palms and road markings, and was told exactly where the ground plane sat so nobody added a second floor on top of it.
• Play owned the player, on-foot movement, vehicle entry, driving, collision, the companion, police heat, the mission and the failure states.
• Presentation owned the title screen, HUD, minimap, result screens, post-processing and sound, and was allowed to adapt the PS1 shader it had pulled in.

The root agent kept the main scene and the shared module for itself and told the others to ask before touching them. Each sub-agent had to write a README for the assets it authored and report its controls, its public API and what it had actually verified. Only the root published.

That is the same division of labour a small studio would use. The difference is that the whole studio was one prompt session.

Sixteen commits in two days

ZeroMind keeps the world's commit history, so the timeline is a matter of record. The first commit landed at 14:16 on 8 September with 658 files: free roam, pedestrians, hospital respawn and police collisions already in. By 22:32 the same evening the world had 2,921 files. The final commit, on the afternoon of 9 September, brought it to 3,366.

The commit messages read like a changelog from a human team:

• Fix locomotion and expand Vice City with connected districts and playable jobs
• Rebuild city surfaces and polish combat, police, traffic and live UI
• Control texture detail in PSX shader and guard visual reload lifecycle
• Add Trailer 1 PS1 cinematic with 47 shots and menu playback
• Add toggleable PS1 trailer soundtrack with synchronized transport
• Rebuild apartment pickup from observed footage with playable transitions and dedicated performances

The trailer commit alone added 1,365 files. Each of its 47 shots is authored content: cameras, actors, props, timing, all driven by the engine's own cutscene runtime, which handled letterboxing, fades, input capture and restoring control afterwards.

The prompts that moved the build

Between the first commit and the last, the agent's notes record a handful of interventions from the person driving it. They are worth reading because they are short.

User requested an actual open world, pedestrians, hospital respawns, and correction of overlapping police cars.

That one prompt turned a single scripted mission into free roam. The city grew from a rectangular road ring to six avenues and six cross streets across 573 by 674 metres, with 87 building footprints, 28 pedestrians on continuous routes and 12 civilian cars you can take while they are stopped.

The previous version was overstated as playable. The user's reported pedestrian resets were real animation root motion, not route resets.

This is the correction that matters most for anyone doing this themselves. The person playing the build saw pedestrians snapping back along their routes and reported it. The agent had called the version playable. It went back, sampled the walk animation, found the clip was moving the character's root 2.4 metres and then looping, and rebuilt the locomotion as an in-place blend driven by actual speed. The fix is in the world. So is the admission.

The final round of prompts supplied reference material rather than instructions: a link to the extended gameplay preview, then a low-frame-rate local copy of it, then a transcript. The agent rendered the video onto a temporary screen inside the engine, reviewed frames, built a shot-by-shot checklist of the apartment sequence against what it had first implemented, and rebuilt that mission with the door, the elevator, the phone call and the red duffel bag in the right order. The subtitled conversations in that mission come from the transcript. The agent notes which lines it left out as ambiguous.

What the engine did so the agent did not have to

Almost none of the build is engine work. The agent wrote Luau scripts, generated textures and models, and authored scenes. The Zero engine supplied the rest:

• The PS1 look is one shader, pulled from another published world (Deata's COINFALL) and adapted: clip-space vertex snapping, affine texture interpolation, nearest sampling, quantised vertex lighting and fog, with a post pass for ordered dithering and 5-bit colour. The whole city, every person and every car use it.
• Reuse from the catalogue. The driving input map, the vehicle materials and the character rigs came from other worlds on ZeroMind. The world's manifest shows content from four other worlds inside it. The agent pulled what existed and wrote what did not.
• Cutscenes, UI and audio are engine runtimes. The 47-shot trailer, the 320 by 240 interactive menus, the radar drawn as a single clipped SVG, and the two soundtracks with synchronised transport are configuration on top of them.
• Validation and publishing. Every commit went through world validation. The final one reports zero errors and zero placeholders. The agent could not publish something the engine considered broken, which is why the world boots.
• Version history. All sixteen commits are visible on the world page. The world has been forked three times already.

Where it is still rough

The agent's audit files are unusually candid, and the world is better for it. From the last audit:

• Mission completions were verified with controlled fixtures and real inputs, but nobody has recorded an uninterrupted start-to-finish campaign run.
• Police tactics and the three cover gunmen are simple AI.
• Character poses, weapon hand alignment and repetitive architecture fall short of a finished game.
• Frame times were spot-read at 23 to 30 milliseconds on one machine, not benchmarked.

If you play it and something breaks, that is the state of a two-day build, honestly labelled. It is also a fork away from being your problem to fix.

Do it yourself

The recipe is not specific to this game. Name the genre in one sentence. Name the characters and the place so the agent cannot drift into generic. Write one paragraph of art direction and have reference images generated from it before any 3D work starts. Split the build into owned areas with boundaries. Then play it, and say plainly what is wrong.

To start, connect the coding agent you already use to the engine (here is how), or open Create and describe the world in the browser. The walkthrough on building a multiplayer 3D world in 15 minutes covers the first session. Everything published lands in the worlds catalogue, where the demake lives now.

Play Grand Theft Auto VI, PS1 Demake. Then fork it.