How Frederik Built Six Games with Claude Code and Codex
Built by ✌︎ frederik ✌︎ · · Source: X
Result: Six games built in a few months. Bolt Pit was cut from 20.5 MB to 12.4 MB and submitted to CrazyGames, where it is waiting for review. 2048 Meadow was rejected by the App Store and the parking game was deleted after one evening.
What they built
✌︎ frederik ✌︎ says they have been a designer for more than 25 years and had never made a game. Over a few months they built six with Claude Code and Codex: Bolt Pit (a twin-stick arena shooter in Three.js and Unity), Rush Hour Sort, a Klondike Solitaire, 2048 Meadow, a browser block puzzle called Yumie Blocks, and a Godot parking game they deleted after one evening.
Their main point is that getting a game to run was quick. Making it fun and getting people to play it took much longer.
How it works
- One subscription, two agents. They pay for one Pro plan and switch between Claude Code and Codex depending on which is better that month. Project notes are plain markdown files, so either agent can pick up where the last session stopped.
- One session controls the tools. Through MCP servers and the command line, the agent controls Unity, Blender, Tripo3D for 3D models and the App Store Connect tools. Concept art comes from ChatGPT Images.
- Bought framework, free assets. For the Unity version of Bolt Pit they bought TopDown Engine, so Claude Code connects existing movement, weapons and enemy AI. Kenney's free asset kits filled gaps in later levels.
- Reference clips for game feel. When fights felt flat, they gave the agent a clip from a game in the style of Brotato and asked what made the hits feel good. It added critical hits, hit flashes, knockback and bigger damage numbers. The first pass went too far, with bullet trails peaking near 2,300 particles, so they cut it back.
- Agent tests plus human play. They say Bolt Pit's suite grew from 118 to 137 passing tests, with 98 automated HUD checks and 752 menu checks. Those show the game works, not that it is fun, so they also play on their iPhone.
- Portal build. Preparing the Three.js version of Bolt Pit for CrazyGames took an afternoon with Claude Code: a separate static build, relative asset paths, removing unused files and a layout fix for landscape phones.
The result
According to the author, the Bolt Pit build dropped from 20.5 MB to 12.4 MB, under the 20 MB limit for the CrazyGames mobile homepage, and the game is waiting for review. 2048 Meadow was rejected by the App Store and is being fixed. A clip of Bolt Pit in Unity reached about 19,000 views without promotion. They expect most of these games to go nowhere and treat each prototype as a cheap test.
How to build something similar
- Start with a game you already like playing, and play each prototype yourself before adding more systems. Passing tests will not tell you if it is fun.
- Keep project notes in markdown files inside the repo so you can switch agents without losing context.
- Show the agent a short clip of a game that feels good and ask it to explain why, then apply the changes one at a time and remove what is too much.
- If you want players, plan distribution early. For a web portal, ask the agent for a static build with relative paths and check the file size limit before you submit.
The original post
I've been a designer for more than 25 years and had never made a game. In the last few months I've built six with Claude Code and Codex. Getting them to run was quick. Making them fun and getting anyone to play them took much longer. https://x.com/i/article/2110280608739909632 I've worked as a designer for more than 25 years, and until recently I had never made a game. Over the last few months I've built six of them with Claude Code and Codex: Bolt Pit, a twin-stick arena shooter where a small robot survives six waves and fights a boss called the Scrap King. There's a Three.js version for the browser and a Unity version. Rush Hour Sort, where you drag cars into the right lane before they reach the wrong exit. Unity. Solitaire, a calm Klondike game in Unity, built because it's a game I actually play. 2048 Meadow, which I submitted to the App Store and which got rejected. I'm fixing it. Yumie Blocks, a block puzzle that runs in the browser. A parking game in Godot that I deleted after one evening. Getting each game to run was the quick part. Making them fun and getting anyone to play them is taking much longer, and that's what most of this article is about. My setup I use Claude Code or Codex, whichever model is better that month. I pay for one Pro subscription and switch it when the other one pulls ahead. My project notes are plain markdown files, so either agent can read them and pick up where the last session stopped. Through MCP servers and command-line access, the same agent session controls Unity, Blender, Tripo3D for 3D models, and the App Store Connect tools I use to ship. For concept art I use ChatGPT Images. When I moved Bolt Pit from Three.js to Unity, I bought TopDown Engine, a Unity framework that already handles movement, weapons, collisions, health and enemy AI. Claude Code connects those pieces instead of writing a physics system from scratch, and I spend my sessions on the parts that are specific to Bolt Pit. For the later levels I filled gaps with Kenney's free asset kits. Find the fun before anything else An agent will build you a game where every system works and nothing is fun to play. My parking game is the example. Cars arrived, you drew a route to the matching parking bay, and two cars crashing cost you a life. It had combos up to 5x, confetti, sound effects and a saved high score, and all of it worked. Then I played it for a few minutes. When the agent asked whether the problem was that nothing was happening or that drawing routes wasn't satisfying, I answered "both". The cars sat still, most routes were obvious and there were no near misses, so I deleted the project that evening. 2048 had a smaller version of the same problem. It worked fine, but I didn't want to keep playing it, and I need to play a game hundreds of times while I build the progression system around it. So I switched to Solitaire, which I already play. The most useful technique I found came from Bolt Pit. Fights felt flat, so I gave the agent a clip from a game in the style of Brotato, a popular arena shooter, and asked it what made the hits feel good. It added critical hits that do double damage, a white flash with a small squash and knockback on every hit, bigger damage numbers, rings on hits and kills, and a short pause after each cleared wave before the upgrade screen appears. A second clip helped with progression. Every kill now gives experience points, so a run offers about 13 upgrade choices instead of 5, and builds end up looking different from each other. Here's the prompt I use for that: The first version of that pass went too far. The feedback was that the game "looks insane after a few seconds", and the bullet trails alone peaked near 2,300 particles. I turned the trails off, cut some of the kill effects and drew bullets smaller while keeping their hit size the same. Art when you can't draw The first version of Rush Hour Sort used shapes drawn in code, and it looked flat next to the concept image. I rebuilt t
Want to build something like this?
Join the Shipping Skool community free. Paid members get live bootcamp calls every week.
Join Shipping Skool — Free