Most people scrolling X right now believe game studios are toast. They see a scrappy one-level demo, slap “gaming is cooked” on it, and move on. This creator did the opposite. He actually built a full 3D game with AI in a few hours, then used it to argue that developers are more safe than ever.
The person behind this experiment is Matt Wolfe, the AI creator known as mreflow. A year ago he vibe coded a simple browser game called The Librarian. This time he wanted to see how far the tools had come in exactly one year, so he built a sequel and documented every messy step. I was genuinely surprised by where he landed, because the takeaway is not the one the hype crowd keeps repeating.
🎮 The key idea
The creator set out to build The Librarian 2, a 3D roguelite where you run around a chaotic library grabbing books that kids keep knocking off shelves. The chaos meter climbs, and when it hits 100 percent, you lose. He wanted procedurally generated levels, boss enemies, natural disasters, meta upgrades that carry between runs, and the occasional “broken run” where your character gets so strong the game feels effortless.
Here is the wild part. The final build came out to 10,400 lines of code across 33 modules with zero asset files. Every texture, character, sound, and piece of music was generated in code. No downloaded art. No sound library. That alone shows how far these models have jumped.
Old way vs new way
This is where the contrast really lands. A real AAA studio spends months before a single line of code gets written. The original poster walked through what that looks like:
- Game directors, creative directors, and design directors all pulling in different directions
- Art teams building mood boards, concept art, and color palettes
- Deep work on story arcs, character arcs, and core pillars
- Studying competing games to find what makes them tick
The new way skips almost all of that. The creator admits he blew past most of the story and narrative planning. Instead he opened a fresh project, pointed the AI at his old Librarian repo for reference, and rambled his whole vision into it using voice dictation. Then he let it run.
He also split the work across two coding models, which I think is the smartest move in the whole video:
- He used Claude Code with the Opus 5 model for the initial build, since that model has been strongest at producing good looking games
- Then he switched to Codex with GPT-5.6 Sol for the fine-tuning and balancing
What blew me away is that Claude Code opened the game in its own browser, took screenshots, tested itself, and fixed what it found. It even rebalanced the difficulty on its own, noticing that bosses were dominating the chaos meter and adding patience timers to fix it.
🛠️ How he actually pulled it off
If you want to try something similar, here is the workflow the creator followed, step by step:
- Reference the past. He linked the new project to his old GitHub repo so the AI understood the original game loop.
- Brain dump the vision. He used voice dictation to rattle off every idea at once, accepting that V1 would be rough.
- Let it run long. The first pass took about an hour and a half of the AI building, testing, and self-correcting.
- Play, then give targeted feedback. He found reversed controls, a broken menu, and books he could not pick up, then fed those exact bugs back in.
- Back everything up. He pushed to GitHub so he could roll back if a change went off the rails.
- Write a handoff file. He had the AI create a detailed markdown design doc so the second model could pick up right where the first left off.
- Switch models to polish. In Codex, he asked GPT-5.6 to read that doc and suggest improvements before balancing the difficulty.
The result looked genuinely fun. There were tutorials, sprint mechanics, a shockwave move that flings books back to shelves, an after-school rush of kids, earthquakes, tornadoes, a bully named Braden you chase on a mini map, and even a grocery store level where a banana makes you slip and a mushroom makes you grow like Mario.
The honest verdict
Here is what I appreciated most. The creator did not oversell it. When he played with his producer, the balancing was clearly off. The bully was impossibly fast. The tornado barely did damage. The levels mostly differed by color. He quoted the CEO of Take-Two, who once said coding was never the bottleneck in making great games. The hard part is all the tiny judgment calls about what to add and what to leave out.
The original poster flatly rejected the idea that people will stop buying games to vibe code their own. He compared it to movies. It is easier to watch one than to create one, and most of us would rather someone else make the creative decisions and hand us a polished experience. He said he has zero desire to stop buying games, and that he rarely even replays the ones he builds himself.
My honest reaction is that this is the most balanced take I have seen on the whole “is X cooked” debate. AI makes development faster, not creativity free.
Watch the full video to see the game come together, the live playtest chaos, and where you can play the finished build yourself.