ONE PROMPT. 55,000 LINES.

Someone handed Claude 5 one prompt, walked away for a few hours and came back to a playable Call of Duty clone. Fifty-five thousand lines of code, every texture, every sound, every animation generated in code from scratch. It's sitting at almost four million views and so many people called it fake that it actually picked up a community note for a while. And if you're new here, I'm Kevin. I break down what's happening in AI and how to actually build with it. If that's your thing, hit follow. So the guy is Matt Shumer, the same guy who wrote the viral Something Big Is Happening With AI post, and he answered the doubters by open sourcing the entire repo and posting the prompt. And that prompt is 150 words, and we have to go over it. Here's the gist of what the prompt is: Build a first-person shooter at the level of the newest Call of Duty. Fan out sub-agents. Have a separate sub-agent act as a really harsh critic. Compare ours against real Call of Duty side by side, blind, and pick the better one. If ours loses, keep going. Loop until it's perfect. That's essentially the whole prompt. He never gave it an architecture, never said how to set up the work or how to do this well or how many rounds to run. I'll be honest, I do the opposite a lot of times. I write the 3,000-word spec, I name every file, I hand the model the exact route, and then wonder why the model quits on me halfway through or only spends about 10 minutes working. And that's choking the model. And Anthropic's own prompting guidance says the same thing. Give these newer models the destination and let them pick the road. Matt calls this the gauntlet loop, and three pieces make it run. Number one, a real bar. He used actual Call of Duty screenshots so the agent could see and set its work down next to them and then lose. A bar has to have something that it can lose to. Number two, blind critic. The builder never grades itself. A fresh critic comes in with no memory of the decisions and does a blind A and B and then sends the work back. So it doesn't know what's Matt's and what's Call of Duty. Number three, no round count. Not three rounds, not five rounds, unlimited rounds. He was actually the one who stopped it. While the agents were still looping to make it better, Matt cut them off. And the game is the least interesting part of all of this. That same loop runs on a SaaS tool you're sick of paying 40 bucks a month for. So here's what to do. Pick a tool you don't want to pay for anymore or that you want to rebuild yourself. Screenshot it. Tell Claude Code to build your version. Use those screenshots as the bar, spawn separate critics with fresh context and give it the no stopping point. Then go to bed and see what it does.