--- title: How to Build and Host a Game With AI (Bloodbench Workflow) url: "https://goremy.ai/blog/articles/how-to-build-game-with-ai-codex-claude-hostinger-vps" author: Luis Chavez-Mattos editor: Luis Chavez-Mattos editor_title: Director, Product published: "2026-10-29T08:40:06.093Z" updated: "2026-10-29T08:40:06.093Z" topic: AI Coding Tools description: How one builder used Claude for research and art, Codex for code, AI testing and an Ubuntu VPS to ship Bloodbench, an LLM benchmark game. --- # How to Build and Host a Game With AI (Bloodbench Workflow) > How one builder used Claude for research and art, Codex for code, AI testing and an Ubuntu VPS to ship Bloodbench, an LLM benchmark game. Edited by Luis Chavez-Mattos, Director, Product ยท Published 2026-10-29 ## How can you build and host a full game with AI? You split the job by tool and act as director. Use a chatbot's deep research mode to build a reference wiki, have a design model produce the art, hand the engine and rules to a coding agent in its desktop app, let AI run thousands of test games, then deploy to a rented Ubuntu server. That is how Wes Roth built Bloodbench, a game where large language models manage fantasy football teams through a full league. ## TL;DR - Bloodbench is a **benchmark game for LLMs**: models manage a fantasy sports team across a league in a game with little training data, so they must learn the rules from scratch. - The builder says he acted as **director, not coder**: his time went to world building, teams and rules, while AI handled code, graphics and most testing. - The workflow ran in order: **deep research into an LLM wiki**, visual style with Claude, engine in Codex, AI playtesting, then hosting on a VPS. - The core build took roughly **10 hours of running time**. Balancing and extra features took much longer, by choice. - A badly structured approach to coding **burned through multiple weeks of Codex quota** and forced upgrades and resets. - Desktop apps from the big labs now make the **command line optional** for agentic coding. ## What is Bloodbench? Bloodbench is a personal project inspired by the Blood Bowl corner of the tabletop fantasy world. Orcs, elves, dwarves and humans play a violent version of football that they barely understand, so it plays more like a battle. The builder says it has no connection to any football league or game franchise. The point is evaluation. Can a language model run a team through an entire league, improve over time, level up players and pick a sensible strategy? Some teams pass, some rush, and some mostly chase casualties. Because the game is obscure, models cannot lean on memorized knowledge. Dice rolls drive many actions, so good play means positioning and understanding odds. Players develop across seasons and can die. The build includes: - Playable single-player mode with team building, opponent choice and a legal starting formation. - Computer-controlled teams to compete against LLM opponents. - A way for outside language models to connect and play matches against each other. - Position stats, starting skills, leveling progression and uniforms, with animations planned for idling, attacking and throwing. ## How does the build workflow run, step by step? 1. **Research.** Open Claude, enable research from the plus button in the normal chat window, and ask for an LLM wiki on your topic. List exactly what you want to understand, and separate official material from community-made material. 2. **Visual style.** Iterate with Claude on the look until it stops repeating itself. 3. **Engine.** Hand the rules, teams, players, AI opponents and interface to a coding agent. 4. **Testing.** Have AI play thousands of games, find issues and tweak the balance. 5. **Hosting.** Put the finished project on a server so anyone can reach it. ## What is an LLM wiki and why start there? An LLM wiki is a set of cross-linked text documents stored locally, built by a research agent from many sources. In the walkthrough, Claude assumed a Karpathy-style wiki structure when asked for one on tabletop fantasy games. The reasoning is simple. A coding agent writes better rules code when it works from a organized reference than from its own vague recall. The creator's codex summary even described researching the rules before translating them into code. Separating official, trademarked material from community-created material also helps you decide what your own project should borrow and what it should invent. ## How did Claude design the art? The first round of art from Claude was repetitive: the same idea with slight variations. The builder pushed back and told it to get more creative. The result included a Metal Slug style retro pixel art direction, which he picked because it fit the game. His broader point on visuals: for a strategy game, a simple readable board beats fancy graphics. He built an isometric view and experimented with 3D, but says clarity matters more. Animations came later. He notes that models handle much of that work alone, and that connectors such as MCP servers can fill gaps where they cannot. ## What did Codex build, and what does the desktop app change? The builder used a model he called GPT-6 Astra inside Codex to write the entire engine: rules, players, teams, opponent logic, the human-playable interface, the LLM connection layer, and the version history, notes and documentation along the way. He argues the command line is no longer the barrier it was. Codex from OpenAI and Claude Code from Anthropic now ship as desktop applications that he describes as at least as capable as the CLI and much easier to use. The CLI remains useful, but you do not need to start there. He also showed ChatGPT taking over his computer through computer use to set up a live game for the demo, which is a reminder that these tools can drive a desktop as well as write code. ## How do you avoid burning through your coding quota? This was the costly lesson. The builder says one approach to prompting the coding model consumed several weeks of his Codex allowance. He needed multiple quota resets and an account upgrade. He describes a right way and a wrong way, and says he would now take a different approach if starting over. The portion of the source available here does not spell out that approach, so treat the following as general practice, not his exact method: - Do research and planning in a separate chatbot, as he did with Claude, so the coding agent's quota goes to code. - Give the agent a tight, well-specified task rather than an open-ended one. - Keep notes and documentation in the repository so each session starts informed instead of rediscovering context. - Check quota consumption early in a project, not after a week. He also split work across tools partly to hedge against running out of Codex capacity, and notes that everyone has a favorite chatbot for particular tasks. ## How does AI testing work for a game? Because the game logic is deterministic code with dice, the AI can run thousands of matches, surface bugs and balance problems, then fix them. The builder says most testing was done this way. Human time went to balancing and design choices he enjoyed, not to repetitive checking. ## How do you host the project on a VPS? Anything people need to reach, such as a website, a server or an MCP, needs a home that stays online. The builder used Hostinger, which sponsored that segment of the video. Take the endorsement as his, and compare providers yourself. The plan shown was KVM2: | Spec | Value | |---|---| | Virtual CPU cores | 2 | | RAM | 8 GB | | Storage | 100 GB NVMe | | Bandwidth | 8 TB | He picked a 24-month term, a United States server location for latency, and a plain Ubuntu operating system with no preinstalled apps. He chose Ubuntu out of personal preference and says to pick your own favorite. The promotional price was quoted in the video, but the transcript is unclear, so check current pricing directly. After provisioning, the generic path is to connect over SSH, copy the project over, run it as a service and put a web server in front. The walkthrough text available here stops before those steps, so follow your host's documentation. ## Is building a game this way realistic for you? Partly. The 10-hour figure covers the agent's running time for the core build, not the total effort. The builder spent far more time on balance and design, and the UI was still clunky at demo time. Treat it as proof that a solo builder can reach a playable, hosted prototype fast, not a polished product. The advantage comes from clear direction: knowing what game you want, what rules matter and what a good result looks like. ## Frequently Asked Questions ### Do I need the command line to use Codex or Claude Code? No. Both labs now offer desktop apps that the builder describes as at least as powerful as the CLI and easier to use. The CLI is still available if you want it. ### What is an LLM wiki? It is a collection of cross-linked text documents, stored locally, that a research agent compiles from many sources. It serves as a structured reference for later coding and design work. ### How long did Bloodbench take to build? The builder puts the core running time at about 10 hours. He spent much more time on balancing, extra features and design, which he enjoyed. ### Why did the builder run out of Codex quota? He says one way of directing the coding model consumed several weeks of allowance, requiring resets and an upgrade. He did not detail the fix in the material available, but separating research from coding and scoping tasks tightly are sensible safeguards. ### What server specs did he use for hosting? A KVM2 VPS with 2 vCPU cores, 8 GB of RAM, 100 GB of NVMe storage and 8 TB of bandwidth, running plain Ubuntu in a US location.