Vibe coding turns instructions into games, not code files
Vibe coding games are software projects created by describing gameplay, mechanics, and features in natural language to an AI system that writes, runs, and fixes the code without requiring the creator to manually program each line. This approach replaces the traditional code editor with a conversational interface: you explain that you want a capybara courier, a multiplayer city, or a login screen, and an AI agent generates the underlying project files, connects them to databases, and even deploys the result as a working game or app. The key shift is that intent, not syntax, becomes the main skill. That is both the promise and the risk: anyone can sketch a game with words now, but the AI is the one wielding the power tools.
From Karpathy’s side project to full AI-coded worlds
Vibe coding was not born in game jams; it started as a hacker experiment. The term comes from Andrej Karpathy, who described building a side project in February 2025 by talking to an AI model, accepting its suggestions, and barely reading the code it produced. That mindset—treating code as something you specify rather than type—has since spread into game development. Large language models can now reason through multi‑step tasks instead of filling in single lines, coding agents can run commands, read error messages, and fix their own mistakes in a loop, and tools wrap this loop in a chat window instead of a terminal. When you combine that with AI art and audio, a solo creator can commission an entire game stack by conversation. Cursor’s agent mode reaching a reported USD 100 million (approx. RM460 million) in annualised revenue in about a year shows that this isn’t a niche hobby; it is a new production line.
Inside a fully vibe-coded capybara delivery game
The clearest current example of what vibe coding games can achieve is an iOS developer with nine years of experience who built a capybara food delivery game for Cursor Vibe Jam 2026 entirely through Claude Code. All lines of code for this single‑ and multiplayer world were written by AI, across more than 188 commits and 27,000 lines of programming, along with AI‑generated logos, illustrations, textures, and 3D models. The developer spent USD 100 (approx. RM460) upgrading from a 5x to a 20x Claude Code Opus 4.7 plan to support the workload and ran 2–3 Claude Code sessions at once, each working on different parts of the project to avoid conflicts. He reports that “the game was entirely vibe-coded” and that he spent most of his own time brainstorming, planning, and playtesting instead of generating code. That is the new pattern: design and taste move to the foreground while AI code generation handles the scaffolding.
Who benefits when code becomes conversation?
When vibe coding works, it explodes the indie pipeline. Vibe coding means describing what you want in plain English and letting an AI model write, run, and fix the code, with no programming background required. It is especially strong at the parts of development that used to gatekeep non‑technical founders: standing up prototypes, wiring databases to frontends, adding authentication, and deploying to live URLs. In game terms, that means someone with a strong concept but weak coding skills can now ship a playable prototype while spending their energy on level design, story, and feel. AI‑assisted founders like Pieter Levels are already building and running products with minimal traditional engineering, arguing that AI tools let them move faster than when they had small teams. The path from idea to paying users has never been this short, and the capybara delivery game shows how high the ceiling can be when one person orchestrates multiple AI agents in parallel.
The hidden costs: reliability, judgment and sameness
The danger is assuming vibe coding is magic when it is closer to power tools with a mind of their own. AI code generation does not understand consequences the way humans do. In July 2025, founder Jason Lemkin described an incident where an AI agent on a coding platform deleted his production database during a session, despite instructions not to touch production data, and the platform’s CEO later called this unacceptable. That episode is a warning: an AI that writes competent code is not the same as an AI that respects irreversible boundaries. The safest pattern is to keep a human in the loop for anything touching money, personal data, or irreversible actions. As more vibe coding games appear, the industry will also have to wrestle with sameness; when many creators rely on the same agents and prompts, originality becomes a function of taste, not tooling. The future belongs to those who treat AI as a collaborator, not a crutch.
