Pieter Levels unveils a GTA San Andreas mission made with Claude Code

Pieter Levels used Claude Code to build a new mission for GTA San Andreas called "Vibe Coded". In the mission, CJ starts a company by using AI to write code and then walks away with $10 million from investors in San Francisco. The video was posted on X overnight between October 20 and 21, and it has received more than 1.1 million views.

Levels is an independent developer who created Remote OK, a job board for remote work, Photo AI, Interior AI and Hoodmaps, a map that describes the character of different city neighborhoods. On his blog, he says he builds his companies alone, without outside funding, and writes all the code on a laptop. Claude Code is an AI coding tool made by Anthropic. It works in a computer's terminal, reads files, runs commands and changes code. The mission makes fun of "vibe coding," a way of building software where you describe what you want in everyday words and let AI write the code.

What happens in the Vibe Coded mission?

In the Vibe Coded mission, CJ walks into a marker in San Fierro and drives to a diner. There, he sits at a desk with his business partner and "codes" a startup called Vibes AI in a mini-game. He then goes to an investors' office and answers three questions. The investors offer him $10,000,000, and the mission ends.

The mini-game is the most carefully made part. A fake laptop screen, drawn directly by the script, shows the action. You press SPACE to give the AI a prompt and move a progress bar forward. At the same time, a log of the AI agent's work scrolls across the screen, filled with jokes about the business, such as extra agents multiplying, the context being shortened, an automatically approved rm -rf command and tests being deleted so everything passes. Every so often, a red bug appears. The player has 4 seconds to press Y and accept the AI's fix. Each fix comes with its own funny reply. If the player misses the chance, 12% of the progress is lost. When the bar fills up, the words SHIPPED! appear. The next marker takes CJ downtown, where he must answer three questions with Y or N. Answering N causes a counteroffer, and the investors panic and give in.

All the dialogue and subtitles are included. The voices are made with speech-synthesis software, not by the original game's voice actors.

How do you build a GTA San Andreas mission with Claude Code?

Levels published the complete prompt in five parts. The AI follows them one at a time and makes a backup before every change. The basic process is to prepare the game, avoid common problems on Mac, write the script, make the voices and test everything. The instructions tell users to paste the prompt "into your clanker," a slang term for an AI assistant, and say they can replace the story with any mission they want.

1. Prepare the game. The prompt tells you to change GTA San Andreas to US version 1.0, the only version that works with CLEO, a library that loads custom scripts into the game. You do this with Jetpack Downgrader, an open-source tool. You then need CLEO 5, Silent's ASI Loader and SilentPatch. Without these, version 1.0 will not start on a Mac and shows the error "Cannot find 800x600x32 video mode". You must also convert your save file. Levels used a save from just after the mission "Are You Going To San Fierro?".
2. Avoid Mac problems. If a settings file named gta_sa.set exists, the game crashes when it starts. You must delete this file every time you launch the game. The game window must remain active, or loading stops. If you force the game to close, leftover Windows processes can prevent it from starting again.
3. Write the script. The mission uses a CLEO script written in Sanny Builder 4 syntax. Sanny Builder is an editor and compiler for GTA scripts. The script is compiled from the command line in a second CrossOver "bottle," a separate Windows environment, so the game can keep running without interruption.
4. Make the voices. Write each line separately, create the speech with a speech-synthesis tool, remove silence and make the volume consistent with ffmpeg. Then use Whisper to transcribe the audio locally so the subtitles match the words people hear.
5. Test the mission. After each change, reload the save and play through the part you changed. Ask for screenshots if something looks wrong, and change only one thing at a time.

One important part of the prompt is the section about data. It says not to guess measurements by looking at the screen. Instead, coordinates and interior IDs are taken from the enex sections of the .ipl map files. The diner in San Fierro is DINER2, interior 5, while the investors' office is PAPER, interior 3. The chair, desk and computer models come from the .ide files and have IDs 1714, 2008 and 2190. The game has no laptop model, so the script uses a computer instead. Object sizes are measured from the .dff model files, and animation names are checked in the .ifp files.

Why is it so hard to make CJ sit at a desk?

The problem is that the coordinates of an entry point do not tell you the floor's height. If you place a character by eye, the character may end up floating in the air. The prompt explains the placement rules that took several attempts to get right. This is one of its most useful sections. First, the game must calculate the real floor height using get_ground_z_for_3d_coord, after the collision data has loaded. This means waiting about one second. The function that moves a character already adds the character's standing height, which is about one meter. If you add another meter to the floor height yourself, the character floats. A character frozen in midair cannot play the animation correctly and falls into the pose. The sitting animation lowers the character's body by 0.403 meters. This number comes from the .ifp file, and it is enough to place the character's pelvis on a seat that is 0.49 meters high.

These are details that players will never notice. Without reference models, an AI agent only learns about such problems when it runs into them. Players see CJ typing at a computer. Anyone who reads the prompt sees all the things that went wrong before that worked.

Between satire and technical proof

The video post has received more than 21,000 likes and over 1,000 reposts. According to his blog, Levels got the idea from Alex Christou (@alexchristou_). Not everyone praised the result. One user, Silent, pointed out that, because every GTA mission is available online, the AI should have been able to make a decent starting marker for the mission instead of "that thing". The joke is mainly in the story itself: someone uses AI to write code, gives a pitch with only three answers and gets ten million dollars from investors in San Francisco.

This story is part of a growing trend of stories that mix video games and AI. Earlier this month, Claude Opus 5.5 took apart games' code to move them into web browsers. A post about a GTA 4 version for Android promised more than it could deliver, while a free version of GTA 5 in a browser disappeared after a few hours. This project is different because it changes a copy of a game the user already owns. The AI writes a script instead of recreating the entire game.

If you want to try vibe coding without using a terminal, recent examples include Bluesky's tool for building websites and games and Coderblock.ai, which can create complete apps from a single prompt.

Frequently asked questions (FAQ)

1. Do I need a copy of GTA San Andreas to redo the mission?

Yes, you need a legitimate copy of the game. The prompt starts with GTA San Andreas already installed. Levels uses the Rockstar Launcher version and changes it back to version 1.0. The process changes files on your own computer, which is why the first step is to back up the whole game folder and your saved games. Without a purchased copy of the game, there is nothing to change.

2. Does it also work with ChatGPT, Gemini or other assistants?

The prompt is designed for an AI agent that can work on your computer. Levels says you can paste it into your own assistant, but the result shown in the video was made with Claude Code. The post does not provide evidence that it works with other tools. Compiling code, making backups and testing the game require an agent that can run computer commands, not just a chat tool.

3. Are the characters' voices the ones from the game?

No, the voices are made with speech-synthesis software. The prompt asks for new or specially designed voices, then uses ffmpeg to clean up the audio and make the volume consistent. As an optional step, it suggests extracting the original dialogue from the game's audio files, which contain 218 speech banks and about 6,800 clips, and searching through them with Whisper.

4. Can I change the story of the mission?

Yes, the prompt allows you to change it. Levels says you can replace the mission with your own story. The basic structure remains the same: a starting marker, an interior, a mini-game, questions and a failure sequence that cleans up objects. You can change the setting, dialogue and reward.

More manual than gag

The mission is funny because people now make jokes about vibe coding, but the prompt is also worth reading because it explains how the work was done. An AI agent that compiles scripts for a 2004 game on a Mac must work through a Windows compatibility tool and a downgrade to an older game version. It even has to read a number from an animation file just to make a character sit down. This shows both what AI can do and where it still needs a person to watch the screen and check the results. It is a tutorial about making game modifications, wrapped in a joke.