Pieter Levels Vibe Coded a GTA: San Andreas Mission. The Real Masterpiece Is the Prompt

The most-shared AI project of the week is a 22 year old video game mission

Pieter Levels, the indie developer behind Photo AI, Interior AI and Remote OK, asked Claude Code to build a brand new mission for GTA: San Andreas. In the "Vibe Coded" mission, CJ steps into a marker in San Fierro, drives to a diner, sits down with a cofounder and ships a startup called Vibes AI on a fake laptop, then pitches three answers to a room of investors and walks out with $10,000,000. The video has been watched more than 1 million times (1,082,305 views and 21,337 likes at time of writing), and the full breakdown of how it was made is worth a slow read (Pasquale Pillitteri's writeup is the best one).

Image: Pieter Levels (@levelsio), via levels.io

Most coverage stops at the joke

The hooks write themselves: an AI agent modding a legendary game, a playable parody of San Francisco fundraising, a minigame where you press SPACE to feed prompts into a progress bar while an agent log scrolls jokes about deleted tests and auto-approved rm -rf. The idea came from Alex Christou, and Levels published the entire five part prompt that produced it.

So the internet did what the internet does. 250+ quote posts, reaction clips, and a lot of "AI can do anything now" hot takes. What almost nobody did was read the prompt. That is the actual artifact here, and it is the most instructive piece of AI direction I have seen in weeks.

My take: this prompt is a design spec in disguise

Here is what the prompt actually does, according to the breakdown:

• It forbids eyeballing. Character coordinates are read from the game's .ipl map files, chair and desk models from the .ide files, model sizes measured on the .dff files, animation names checked in the .ifp files. No "put CJ roughly at the diner." Every placement is sourced from source-of-truth data. That is why CJ sits in a chair instead of floating a meter above it.
• It forces one change at a time. After every edit, reload the save, replay the part you touched, request a screenshot when something looks off. That is not a hobbyist workflow. That is a QA loop.
• It builds defensively. Backup before every change, compile in a separate environment so the running game is never disturbed.
• It specifies the weird edge cases. The sitting animation lowers the body by exactly 0.403 meters. The seat is 0.49 meters high. The function that moves a character already adds standing height. Pass "floor plus one" and your character floats.

Sound familiar? It should. That is acceptance criteria, source-of-truth tokens, visual regression testing and a design system doc, applied to a game script. Designers have been writing specs exactly like this for a decade. We just called it boring documentation instead of prompt engineering.

Most people still think the skill gap in AI is "prompting," the magic-words idea that the right incantation unlocks good output. Projects like this show the gap is really direction: knowing what "correct" means before you generate, proving it against real data, and iterating in small verified steps. The people who will do best with AI coding agents are the people who were already rigorous about specs. Designers are massively overrepresented in that group and mostly do not realize it.

What to steal for your next AI build

1. Measure, do not eyeball. Give the agent source-of-truth files, design tokens, real data. Every "roughly here" becomes "floating in midair" somewhere downstream.
2. One change at a time. Batch edits feel faster and make regressions untraceable. Small steps with verification beat one heroic prompt.
3. Ask for screenshots. On a feeling that something is off, the workflow demands visual proof. Your AI build pipeline should too.
4. Write the edge cases down. Heights, offsets, tolerances, the boring numbers. This is where agents fail silently.
5. Keep the backup before every step. Rolling back should cost one command, never a rebuild.

The satire is aimed at all of us

Let's not pretend the joke is subtle: a three-answer pitch walks out with $10M from San Francisco. Ha. But the constructive reading is better: Levels put real craft into mocking hype culture, and the craft is the transferable part. A stunt with a rigorous spec behind it is infinitely more valuable than a vague "AI will change everything" thread with none.

So no, I do not think this is a distraction from serious AI work. I think it is the clearest demo of the year of what serious AI work looks like when the output is fun: a tight spec, trusted source data, tiny verified steps. The mission ends when the progress bar says SHIPPED. Yours should too.

If this made you want to direct an agent through your own product build instead of just prompting and praying, try Base44 and apply the same rules: source-of-truth data, one change at a time, screenshot everything.

Related on NewForm

• Meta Muse Wonder: Frontier AI Now Costs Up to $200 a Month, and That's the Story
• Epic Games Vibe-Coded Its New Launcher. The Real Story Is the "AI First" Order
• Claude Code's Graceful Stop: What the 5-Hour Limit Fix Means for Builders and Designers