A man studies a laptop in a diner overlooking a city skyline and suspension bridge at dusk.
Pieter Levels, the indie developer behind Remote OK and Photo AI, asked Claude Code to build a new mission for the PC version of GTA: San Andreas. The mission is called "Vibe Coded." In it, CJ vibe codes a startup and raises money from San Francisco investors. The joke is fun, but the developer-facing material is the public prompt behind it. That prompt is a detailed record of what it takes to point an AI coding agent at a 2004 game, a Windows-era toolchain and a Mac compatibility layer.

Levels' video post on X is stamped October 20, 2026 and shows 1.1M views. His follow-up post, which contains the full prompt, is stamped October 21, 2026, 4:31 PM. The two dates describe two separate posts. View counts keep changing, so treat them as snapshots. Levels credits Alex Christou as the source of the idea.

A man studies a laptop in a diner overlooking a city skyline and suspension bridge at dusk. What the mission does​

CJ walks into a marker in San Fierro and drives to a diner. He sits at a desk with a co-founder and "vibe codes" a startup called Vibes AI in a minigame. He then drives to an investor office and pitches. The investors put in $10,000,000 and the mission passes. All dialogue is voiced and subtitled.

The prompt adds more detail:

  • Minigame: a fake laptop screen is drawn by the script itself. The player taps SPACE to prompt the AI and raise a progress bar.
  • Agent log: a scrolling log carries jokes about spawning subagents, context compaction, an auto-approved rm -rf and deleting the tests.
  • Bug events: random red bugs give the player 4 seconds to press Y and accept the AI's fix. Failing costs 12% progress.
  • Pitch: three Y/N questions, then the $10M offer. Answering N triggers a counter-offer gag in which the investors panic and cave.
  • Voices: generated with text-to-speech, not the original cast. The prompt does not name a speech service.

The toolchain behind the gag​

This is a modification of a game copy you already own. It is not a browser port or a recreation of the game. The prompt targets the classic PC version running on macOS through CrossOver.

Its main steps:

  1. Back up everything. Copy the full game folder and the save folder before touching anything.
  2. Downgrade to v1.0 US. The prompt says CLEO only works on v1.0 and that the Steam and Rockstar Launcher builds are newer. It names the open-source Jetpack Downgrader and applies its binary patches through CrossOver's Wine command line. It gives checks for the result: a gta3.img of 940,064,768 bytes with CRC32 c8ccf06f, and a main.scm of 3,079,599 bytes.
  3. Install CLEO 5 and SilentPatch. The prompt says SilentPatch is required because v1.0 otherwise refuses to start on a Mac with a "Cannot find 800x600x32 video mode" error.
  4. Use a compatible save. v1.0 uses v1 scripts, so saves from newer versions won't load. The prompt uses a converted save from just after "Are You Going To San Fierro?"
  5. Launch the game directly. Run gta_sa.exe itself, not the Rockstar Launcher, which may undo the downgrade.

CLEO's own documentation backs up the version boundary. Its requirements list GTA: San Andreas PC v1.0 US along with Silent's ASI Loader or an equivalent. That makes v1.0 a requirement of this CLEO route. It is not a rule for all GTA modding. The same documentation says a compiled CLEO script is a .cs file dropped into the CLEO folder. That matches how the prompt installs the mission.

Where the agent needed hard facts​

The prompt is firm that nothing should be guessed. It tells the agent to pull real data from the game files:

  • .ipl map files: interior entry coordinates and IDs. The prompt uses the San Fierro diner (DINER2, interior 5) and the Planning Department office (PAPER, interior 3).
  • .ide files: model IDs. The prompt uses chair 1714, desk 2008 and computer 2190, and notes that the game has no laptop model.
  • .dff models: object sizes, measured from vertex bounds.
  • .ifp files: animation names, checked by parsing the files.

The script is written in Sanny Builder 4 syntax in SA v1.0 SBL mode. It is compiled from the command line through Wine, in a different CrossOver bottle from the running game so a build never disturbs the game. The prompt also warns about two Sanny Builder pitfalls: variable names that clash with built-ins such as "blip" and "fx", and a cap on local variables. CLEO's documentation confirms the cap, listing 32 local variables.

The "why is CJ floating" lesson​

The most instructive section covers placement bugs that took several attempts to fix:

  • Entry-point coordinates are not floor height. The real floor comes from get_ground_z_for_3d_coord, called only after collision has loaded (after load_scene and about a second's wait).
  • set_char_coordinates already adds the ped's standing height of about 1 m. Passing "floor plus one" makes characters float.
  • A character frozen in mid-air plays a falling pose, not its animation, so freeze only once they're on the floor.
  • The sitting animation lowers the body by 0.403 m (read from the .ifp), which lands the hips on a 0.49 m chair seat.

These numbers are the prompt's own reported measurements. I have not checked them against the game archives. They show why such a task needs measured data and in-game testing, and why eyeballing it doesn't work.

Mac and CrossOver gotchas​

The prompt lists several workarounds specific to Levels' setup:

  • Delete gta_sa.set before every launch, because a settings file from a previous session crashes the game at startup. Levels used a small .bat wrapper, and the cost is that in-game settings reset each launch.
  • Keep the game window focused during loading, since loading stops otherwise.
  • Force-quitting leaves orphaned Windows processes that silently block future launches. Kill them or use Force Quit / Restart Bottle.
  • If the camera spins, set the Wine registry value MouseWarpOverride to force under HKCU\Software\Wine\DirectInput.
  • Don't enable Retina mode, which broke launching in this setup.
  • Windowed mode through dxwrapper produced an invisible window, so it was removed.

These are reported fixes for one environment. They are not universal CrossOver guidance.

Testing and voices​

The prompt's testing rules are plain and sensible: work step by step, back up before changes, check the compiler log after every build, reload the save and replay the changed section, request screenshots when positions look wrong, and change one thing at a time.

For audio, it asks for a list of every spoken line, one file per line. The clips are generated with text-to-speech, trimmed and loudness-normalized with ffmpeg and converted to mono mp3. They are then transcribed locally with Whisper so the subtitles match what is said. The clips play through CLEO 5's audio plugin. If a clip fails to load, the subtitle shows without sound instead of crashing the game. The prompt also suggests an optional route to extract the game's own speech banks (218 banks, about 6,800 clips).

Caveats​

  • One recording is the only proof of the mission. The video shows it running. It does not show that another person's setup will produce the same result.
  • The prompt is a distilled log of trial and error. Levels replied on X that the prompt was Claude explaining how to do it "because it was quite a process." So the public text is a record of a long iterative run. It is not a one-shot recipe.
  • Other people say a leaner approach could work. One commenter claimed a stronger model with a good testing framework could work most of this out itself from screenshots and recordings. That is an unverified opinion.
  • Quality critics exist. One modder pointed out that, given how many GTA missions exist online, the AI should have produced a proper mission start marker. Another commenter said it reused the double marker from Flight School.
  • The route is narrow. It targets classic PC files on a Mac with CrossOver. It is not a procedure for console or mobile editions. Nothing in the public evidence shows it works with other AI assistants.
  • It is a modification, not a game. Levels' own page also notes he has a record of shipping quick vibe-coded projects, and the stunt fits that profile.

Takeaway​

For people who write software, the practical lesson is the shape of the workflow. A coding agent can draft and revise a script. But success here depended on a pinned legacy version, compatibility patches, a compiler running in a separate environment and facts read from game assets. It also depended on repeated in-game testing by a human watching the screen. Claude Code is a command-line tool that reads files and runs commands, which suits a compile-and-test loop like this. Even so, the public record doesn't show how much human supervision went into the final result.

 

References

  1. Pieter Levels unveils a GTA San Andreas mission made with Claude Code - Pasquale Pillitteri Pasquale Pillitteri 2026-10-24T16:10:00+00:00
  2. I asked Claude Code to make a new GTA: San Andreas mission where CJ has to vibe code a startup and raise money in SF — @levelsio (Pieter Levels) levels.io
  3. @levelsio (@levelsio): "Here is the prompt to make your own GTA San Andreas mission on MacOS, just paste it in your clanker, you can change the mission of course to your own story: You are helping me build a custom story mission for GTA San Andreas (classic PC version) running on macOS through CrossOver. Work step by step, check each step works before moving on, and back up everything before changing files. GOAL A playable mission called "Vibe Coded". CJ walks into a marker in San Fierro, drives teapawt.org