Vibe-Code Quest: How One Founder Built a Language-Learning Roguelike with Pure AI Magic
Max Harned didn’t set out to become a programmer. He simply wanted a more engaging way to learn Swedish during his downtime—something closer to a favorite roguelike deck-builder than to a traditional language app. Four months later, he and a part-time collaborator have built LanguageMancer, a mobile game created almost entirely through “vibe coding”: natural-language prompts fed to large language models that generate code, art, audio, and even balance suggestions. Their experiment shows how fast AI tooling is collapsing the gap between a playable idea and a polished product—and why “taste” may soon matter more than technical fluency.
From Curiosity to Prototype
The genesis was a simple frustration. Max could sink hours into Slay the Spire or Vampire Survivors but struggled to spend ten minutes in Duolingo. Armed with models like GPT-5, Claude Sonnet, and Gemini 2.5, he asked the code-aware IDE Cursor to scaffold a PhaserJS project that merged card-battle mechanics with vocabulary drills. Within days, the prototype dealt damage, awarded gold, and slipped foreign words into combat decisions. By week four, it boasted two playable characters, 150 cards, a boss rush, and voice-acted pronunciation checks. Max estimates a traditional studio would have needed ten people and 18 months to reach the same milestone.
The Vibe Coding Stack
Cursor sits at the center of the workflow, allowing Max to chat, “Add a screen-shake when the player takes critical damage,” or “Refactor the combat file into smaller modules,” and watch GPT-5 modify JavaScript across hundreds of files. Midjourney and Ideogram generate character concepts; sprite-sheet tools or freelance animators transform the best concepts into idle loops and attack frames. ElevenLabs supplies granular sound effects—an ominous boss laugh, the whoosh of a fireball—while Suno turns a hummed melody into the game’s orchestral theme. Even card balance passes through the models: Max drops the full card list into a GPT-5 “project,” asks for synergy gaps, and iterates until the AI’s suggestions feel fun in play.
Managing Complexity Without Writing Code
Early on, large files and missing dependencies caused whole modules to vanish. The solution was to keep each source file under about 5,000 lines and to insist on reusable systems—one particle engine, one status-effect manager, one shop UI—so the AI stopped rewriting the same logic. Git integration inside Cursor lets him roll back any destructive edit with a natural-language command. When a memory leak crashed older phones, he hired a human developer for a week, but ultimately solved it by asking successive models to isolate assets in RAM. Roughly 60 percent of token spend still goes to discarded attempts, yet monthly AI costs have dropped from $1,500 to around $500 as models improve and context windows expand to a million tokens.
Economics of a Solo AI Studio
All-in, the project has cost about $6,000: model credits, a handful of $30 animations, small play-test budgets, and minor contractor fees. That figure includes three languages—Swedish, French, and English—and enough modularity to add more with a Python script that calls Google Cloud Text-to-Speech for audio lines. Max now plays his own game for fun, logging twenty leisure hours outside formal testing. He plans a soft launch via TestFlight and Google Play’s open-testing track, then hopes to raise seed funding for a mini-studio that can ship multiple AI-authored learning games.
Taste as the New Moat
The technical hurdles are falling quickly. In a year, Max predicts “add a snake with an orchid head” will summon an animated, balanced enemy ready for production without leaving the IDE. The remaining challenge is creative judgment: deciding which of the thousand AI-generated monsters fits the game’s tone, which musical loop captures the garden biome, which difficulty curve keeps players hooked without frustration. As Suno, RetroDiffusion, and future 2D-to-3D tools automate the craft layer, developers will differentiate by intuition—curating, pruning, and directing models the way film directors guide actors and cinematographers.
Where It Goes Next
Max’s immediate goal is data: watch hundreds of users collide with act-one bosses, harvest crash logs automatically, and feed the results back into GPT-5 for nightly tuning. Longer term, he imagines AI testers that play thousands of runs while he sleeps, flagging over-powered streak builds or under-performing rune combinations. If the App Store approves, LanguageMancer could prove that language learning can feel like chasing loot drops, not filling flash-card quotas—and that a two-person AI studio can out-iterate teams ten times larger. In the emerging era of vibe coding, the real scarce resource is no longer engineering muscle; it’s vision, persistence, and, above all, taste.