Casey Muratori Reviews Vibe-Coded Mining Games: AI's Real Promise and Hard Limits in Game Dev

On this episode of TheStandup, the hosts turned a seemingly absurd premise — using AI to rapidly crank out a bunch of "mining games" — into a real-world stress test of Vibe Coding. Legendary game programmer Casey Muratori was brought in as judge and delivered candid critiques of works produced by large language models in anywhere from two weeks to two hours. What looked like playful chaos ended up exposing the true limits of AI-assisted game development.

From Mining to Everything: The "Divergent" Iteration of AI Prototypes

Trash (one of the hosts) admitted his process was completely free-associative. The original prompt was to make a "mining game," but he quickly found pure sand-digging too boring and went off the rails: from a backyard pool and sponge-soaking mechanics, to a kitchen prep game about clearing a plate with chopsticks, then excavators, a Dung Beetle multiplayer brawler, a trailer simulator, a Karate Kid-inspired fly-catching game, and finally a full mini-game collection website.

He described it as "the evolution of my brain over the past two weeks." At their core, all these games share the same mechanic — clear or collect something as fast as possible within a time limit. This explosive rapid-prototyping capacity is AI programming's most obvious value: ideas can be made tangible at nearly zero cost.

Models Are Good at Modeling, Bad at Code

Trash mentioned that while testing Astra (a newly released model), he landed on a key insight: AI performs reasonably well at 3D modeling, but falls significantly short when it comes to writing code. The models he generated had consistent styles, good variety, and were visually passable — but on the code side, things were "very janky."

More critically, there was a performance problem. When Casey's computer nearly crashed from the fan going full blast because the website was stuffed with too many Three.js models, the technical detail sparked a valuable discussion.

Vibe Coding is a concept coined by Andrej Karpathy in early 2025. It refers to a mode of development where you rely entirely on AI to generate code — the developer barely reads, let alone fully understands, the specific implementation. You describe your intent in natural language, repeatedly accept AI suggestions, observe the running result, and adjust your prompts. The whole process feels more like "sensing" and "steering" than traditional writing and debugging. This dramatically lowers the barrier to prototyping and lets people without a professional programming background rapidly materialize ideas — but at the cost of unpredictable code quality, hard-to-trace deep bugs, and maintenance costs that can far exceed initial development. The hosts' experiments in this episode are a textbook example of this pattern.

Three.js is a JavaScript 3D graphics library built on WebGL that lets developers render 3D scenes in the browser without directly touching low-level graphics APIs. While it dramatically lowers the bar for web 3D, its performance bottlenecks are well-documented: the browser's single-threaded JavaScript execution model, the CPU-GPU data transfer overhead of WebGL, and the lack of automatic scene-level optimization all make it easy for frame rates to collapse and CPUs to overheat when stacking a lot of 3D assets. Professional game engines like Unity and Unreal have full LOD (Level of Detail), occlusion culling, and streaming asset loading systems baked in. AI-generated Three.js code typically doesn't implement any of this automatically — which is exactly why Casey's computer fans went into overdrive.

Casey's Professional Take: This Isn't AI's Fault

As a veteran game programmer, Casey was remarkably calm facing a scenario where so many models were bringing down a website. His take is a valuable warning for AI developers:

"Personally I don't feel much about it, because how badly the web handles this kind of thing has always been kind of a given. You open a 3D website and suddenly the fans kick on for no apparent reason. AI is definitely part of that, but a lot of it is just that the platform is bad."

He went further and traced the root cause back to training data: "Because the people before put bad websites together, the AI learned to make bad ones. I'm not going to blame AI for things I've seen a thousand times before."

This is a measured, professional perspective — the quality ceiling of AI-generated code is largely constrained by the quality of the historical code it learned from, as well as the inherent limitations of the underlying platform (like Web 3D). You can't simply blame the model.

Prime's Work: Same AI, Wildly Different Results — Why?

A recurring puzzle throughout the episode was: why does the visual quality of output vary so dramatically between people using the same AI? Prime's DDR-style mining game — where you press arrow keys in time with music to control digging speed — had art that was "unrecognizable, like corrupted visual noise," and the gameplay logic was completely broken. Pressing an arrow key would just keep digging indefinitely, the beat was totally out of sync, and the arrow keys also caused the character to move left and right, directly contradicting the core concept of "digging down."

Prime admitted this "work of art" cost him over $500 in token spend, and it was still a disaster. The group speculated the difference might come down to prompting strategy: TJ used real photos as reference, while Prime described everything in plain English and let the AI generate from scratch, making quality nearly impossible to control. This reflects a real truth about Vibe Coding — output quality depends heavily on the specificity of input and the presence of reference material, not simply the amount of compute thrown at it.

Casey's technical note was sharp too: if Prime had placed the rhythm cue directly below the character (instead of on the left side of the screen) and made a few other basic game design adjustments, this DDR mining game would actually be the one he'd most want to play. "Game Design 101" fundamentals are not something AI will automatically fill in for you.

TJ's Knockout Blow: Narrative Is the Real Weapon

The work that truly won over the room was TJ's. Using Claude Opus in under two hours, he built a satirical narrative game taking aim at Trash's habit of "digging holes" — a Pokémon battle-style interactive story where the protagonist Basura (Spanish for "trash") gets backed into a corner by their partner's "gotcha" evidence: dirty dishes in the sink, cabinet doors left open, the thermostat secretly turned down. Step by step, they dig themselves deeper until there's no escape.

The most striking detail: TJ said he barely prompted it and the AI automatically referenced specific bits like the "dishes" running joke from their earlier podcast episodes. He speculated that their podcast content may have made it into the model's training data. If true, this suggests that public content creators' words are feeding back into AI outputs in unexpected ways.

Casey ultimately gave the award to TJ, with clear reasoning: "If we had a narrative award, it unquestionably goes to TJ." Of all the submissions, TJ's was the only one that could be called a complete game "as presented" — not for its potential, but for what it actually was.

Anthropic's Claude Opus — the model TJ used — is the most capable version in the Claude 3 series, excelling at long-context understanding, narrative generation, and complex reasoning, though API costs are correspondingly higher. TJ's ability to produce a complete narrative in such a short time was partly due to Opus's deep contextual understanding, but it also confirmed a practical pattern: narrative games place far lower demands on code logic than mechanics-driven games. AI's consistency when generating dialogue branches, plot twists, and emotional pacing is noticeably stronger than when generating physics collisions, rhythm synchronization, or other gameplay code requiring precise numerical control. This "play to your strengths" project selection was the hidden factor in TJ's win.

The "Potential" Trap: The Rhetorical Trick at the Heart of Vibe Coding

A recurring joke throughout the episode was Prime's attempt to "work the jury" — asking Casey to judge by "potential" rather than finished quality. Both TJ and Casey saw right through it.

TJ had one particularly sharp line: "These are all Vibe Coding games, and the entire AI hype cycle is about potential. Nobody is talking about what we already have — everyone is talking about what we might have. So judging by potential is completely in the spirit of Vibe Coding."

This half-joking remark hits on a real problem with how AI products are demonstrated today: a carefully edited demo can make any half-finished product look breathtaking, but the gap when you actually get your hands on it is often enormous. Casey's line — "this is what it actually looks like when you play the game from the demo video for real" — is the most clear-eyed footnote to the whole episode.

Key Takeaways for AI Developers

This seemingly chaotic mining game showdown actually captures the real state of Vibe Coding today:

• Prototyping speed is stunning, but completion is the hard barrier. AI can generate a runnable game skeleton in a few hours, but it won't automatically fill in game design fundamentals or the final 10% of polish.
• Narrative and creativity are the human moat. TJ's win came not from technical skill, but from a sharp idea that landed perfectly on a team in-joke.
• Training data sets the output ceiling. Whether it's code quality or content references, the model faithfully reflects what it was trained on.
• Be wary of the "potential narrative." Between a demo's shine and a product's actual playability lies a mountain of invisible engineering debt.