A practical guide to building a complete game with ChatGPT Codex, Claude, and AI agents, covering code, art, testing, and hosting.
What does it actually take to build a game with AI tools today? #
Building a game with AI tools now mostly means directing rather than doing. One creator recently built a complete turn-based strategy game, a fantasy sports simulator inspired by tabletop games like Blood Bowl, using ChatGPT (via its Codex coding agent), Claude for research and art direction, and a hosting provider to put it online. The core game engine, rules, AI opponent logic, and player interface were coded by Codex in roughly 10 hours of active agent work. The human role was world building, direction, and quality control, not writing code or drawing art.
TL;DR #
- Role reversal is the biggest shift: the person building the game acted as director and world-builder, while AI agents handled coding, art generation, and testing.
- ChatGPT’s Codex agent wrote the entire game engine, the rules logic, AI-controlled opponents, the player-facing interface, and even version history and documentation.
- Claude was used for deep research (building what the creator calls an “LLM wiki” of source material) and for generating visual art styles, including a retro pixel-art look inspired by Metal Slug.
- Desktop coding agents like Codex and Claude Code have gotten dramatically easier to use than raw command-line tools, lowering the barrier for non-programmers to direct real software projects.
- Testing at scale was offloaded to AI as well: the system ran thousands of simulated games on its own to surface bugs and balance issues.
- Hosting matters once a project is playable: running a game, bot, or agent 24/7 so other people (or other AI agents) can connect to it generally requires a server rather than a local machine.
- The biggest risk in this kind of project isn’t the AI’s capability, it’s approaching the build in the wrong order, which can burn through weeks of coding-agent quota on rework.
Remy is new. The platform isn't. #
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
How do you start a build like this? #
The process began with research, not code. Before any game logic existed, the creator used Claude’s deep research mode to build what he called an “LLM wiki,” a structured, cross-linked set of reference documents about the tabletop fantasy genre and the specific rules conventions he wanted to borrow from. This step mattered because the game itself (a fantasy take on American football, with orcs, elves, and dwarves battling as much as competing) isn’t a widely documented commercial product. There wasn’t a large pile of existing training data to lean on, so the research phase gave the coding agent a grounded set of rules and terminology to work from before writing a single line of game logic.
This is a pattern worth generalizing: for any AI build involving a niche domain, rules system, or style the model hasn’t seen much of, front- research with a reasoning or research-capable model (Claude’s research mode is one option) gives the coding agent something concrete to translate into code, instead of letting it guess.
Which AI tool handles which part of the job? #
In this build, the division of labor looked like this:
- ChatGPT with Codex did the heavy technical lifting: the full game engine, player and team data structures, the dice-roll and stat systems, AI logic for computer-controlled teams, the interface for human play, and the scaffolding that let other large language models connect in and play matches against each other. Codex also maintained version history and wrote its own documentation as it worked.
- Claude handled research and visual direction. When asked for character and environment art, Claude initially produced fairly generic variations. Pushing it to “get creative” led to a retro, pixel-art, Metal Slug-style look that became the game’s defining visual identity. Claude’s outputs became the base assets that Codex then wired into the actual game interface.
Splitting work this way wasn’t arbitrary. The creator noted a practical reason: Codex was doing so much of the core coding that he worried about running through its usage quota, so he routed research and art tasks to Claude instead. That kind of quota awareness is a real constraint when working with coding agents over many hours, and it’s worth planning for on any similarly ambitious project.
Why did the project take weeks instead of hours the first time around? #
Despite the engine itself coming together in about 10 hours of agent time, the overall project stretched across multiple weeks, partly because of a wrong turn early on in how the coding agent was directed. Approaching the build in the wrong order led to a cycle of rework that burned through multiple weeks’ worth of Codex’s quota allowance, forcing account upgrades and quota resets just to keep going.
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
The lesson isn’t that the AI failed, it’s that sequencing matters. Letting an agent build deep, specific features before the foundational architecture and rules are solid tends to create cascading rework. Nailing down the research and rules structure first, then having the agent build the engine around that foundation, is the more durable order of operations.
How much of the testing and polish was done by AI? #
Once the core game existed, most of the testing wasn’t manual. The system ran thousands of simulated games on its own, using that volume to surface bugs, balance issues, and edge cases that a single human playtester would take far longer to find. The creator’s own time after the initial build went into balancing decisions and feature additions, the parts of the process he described as enjoyable, while the repetitive verification work was handled by the AI loop.
This matches a broader trend in AI-assisted development: agents are increasingly used not just to write code but to generate their own test cases and run them at a scale no human team would attempt manually.
Is hosting still necessary, or can everything run locally? #
If a project is only ever going to run on the builder’s own machine, local hosting is fine. But anything meant to be reachable by other people, or by other AI agents (in this case, letting different language models connect in and play matches against each other) needs to run somewhere persistent and accessible. That generally means a virtual private server rather than a home computer, since a VPS stays on, gets maintained, and is professionally secured rather than relying on home hardware and a personal network connection.
For this build, that meant provisioning a cloud server with enough CPU, memory, and bandwidth to run the game continuously and accept outside connections, choosing an operating system, and picking a server location close to the intended users for lower latency. The specific hosting setup was sponsored in the source video, but the underlying principle applies regardless of provider: once a project has an audience (human or AI), it needs a stable home online.
Frequently Asked Questions #
Do you need to know how to code to build a game this way?
No. In this project, all of the actual coding, from the game engine to the AI opponent logic to the interface, was done by ChatGPT’s Codex agent. The human role was research, direction, art feedback, and balancing decisions, not writing code.
What’s the difference between using a command-line coding tool and a desktop agent like Codex or Claude Code?
Command-line interfaces require typing commands directly and have historically intimidated non-technical users. Desktop applications like ChatGPT’s Codex and Claude’s Claude Code now offer comparable power with a much simpler interface, making them accessible to people without a programming background.
How long does it actually take to build something like this?
The core game engine in this example took about 10 hours of active coding-agent work. The full project took considerably longer because of an early misstep in how the build was sequenced, which led to weeks of rework and quota resets before the creator found a more efficient approach.
Can AI generate custom art styles instead of generic stock-looking images?
Seven tools to build an app. Or just Remy. #
Editor, preview, AI agents, deploy — all in one tab. Nothing to install.
Yes, but it may take pushing the model. In this case, initial AI-generated art looked generic until the creator explicitly asked for more creative, distinct options, which produced a retro pixel-art style that ended up defining the whole game’s look.
Why does a project need online hosting instead of just running locally?
Hosting matters once other people, or other AI systems, need to interact with the project directly. A persistent server keeps the project running continuously, handles security, and allows outside connections in a way a personal computer generally can’t match reliably.