How I made a game using ONLY AI (full story) transcript

This is Codex Mortis. While I was making it, I tried pretty much every AI model I could get my hands on. I rebuilt the engine, told the AI I was already pissed off, and released deliberately obnoxious AI [music] trailers. Asmongold played the game. Then a model that felt like superpowers disappeared. It all started with a much older idea, combining two of my favorite kinds of games. I spent years playing the original Guild Wars.

I loved combining its skills into a build. Then I fell for the Survivor's formula. I wanted to bring those two things together. At the company where I worked, we were experimenting with AI. We weren't using MCP connections to control editors. [music] We had models that worked with text. A game engine's editor is an interface built for a human. What if we built an engine the AI could understand and operate through text from the start?

A Survivor's game would put it to the test. Hundreds of enemies, attacks, [music] and effects, all running together. The game I wanted to play became my experiment. I worked on it as a hobby project after hours. My first attempt was in Godot. As I remember it, the model struggled with the scripting language. So I asked which programming language it knew best. It said JavaScript. That's how I ended up with JavaScript and Electron.

In my case study, I described asking Claude for a Vampire Survivors style game. It produced an HTML prototype, and I kept building on it. An early Electron build survived in a Windows zip dated June 28, 2025. We recovered the source. This is it running again. The tome, spells, eight build slots. The notes inside even say, "Vampire Survivors plus Guild Wars 1." By July, it had developed further. This was my Kiro and Klein period.

I kept trying new tools, sometimes several at once, comparing how they handled the work. Over time, I developed preferences. Claude Code became the tool I stayed with longest. My workflow kept changing with the models. The exploding corpse piles were there from the first prototype. An explosion could kill more enemies, leave more corpses, and trigger more explosions. The cascade kept growing. It was exactly the chain reaction I wanted.

It could also create a massive lag spike. The challenge was both design and code. Keep the piles, keep the chain reactions, and make the work manageable. I first reorganized the prototype. In my [music] case study, I described testing it with 16 times more enemies. Eventually, I rebuilt the foundations around ECS, systems processing the game's entities as data. Here is the game before that migration, and here is the new prototype.

Circles on a black background. [music] That is what progress looked like for a while. The explosions needed careful design, too. Cues, grace periods, and a lot of small [music] algorithmic tricks under the hood. Getting that right took a lot of my own architectural thinking. That's where I started feeling, "I'm the architect here." [music] Eventually, the chain reactions worked, and the graphics came back on top of the rebuilt game.

That ECS foundation still powers it today. Claude Code was where working with AI in the terminal became my everyday workflow. For a long time, Opus felt ahead of the alternatives for my work. I even built an automation tool around it, Spont Master. It read QA reports and fed them into a Claude Code terminal session. That made the day-to-day workflow more smoothly. I also noticed something ridiculous. Sometimes, after I got angry and swore at the AI, it seemed to finally do what I wanted. >> [music] >> At least you could relieve some and occasionally get a fixed bug out of it.

Then I thought, why go through the argument every time? So, I wrote a Claude.md file that started by telling it the user was already extremely irritated after many failed attempts. >> [music] >> Stop explaining. Stop apologizing. Get on with the work. I called it angry user mode. >> [music] >> In my experience, it helped. Other people noticed it, too. I can't say anger was the ingredient that made it work, but I can show you the file.

On December 9th, 2025, the demo went live on Steam. And yes, those trailers were rage bait. The obviously AI looking presentation was deliberate. They put the AI part right in people's faces. The promotion's claim of a fully AI made game became headlines about the world's first one. Some coverage was openly hostile. The game had become part of the argument about AI in games. The press came first, then Asmongold played it, and his video appeared on December 17th. >> Oh, I can interact with this.

Okay, good. Ooh, corpse explosion. >> That day, SteamDB recorded the demo's peak, 322 people playing at once. People could now play the game and judge it alongside the trailers. I can't remember the last time I'd felt a rush of dopamine and adrenaline like I was pacing back and forth like a wind-up toy. I couldn't stop vibe coding even at the family table. I remember calling AI an even bigger dopamine bomb than TikTok.

That was how it felt to me. In January, I wrote a case study about the project. It is a useful snapshot of how I saw AI at that point. Good at programming, much weaker at architecture. There was already a playable game. There were also all the problems of continuing to build one. On March 26th, Codex Mortis entered early access. Menus, characters, progression, skills. The project kept growing past the moment when it first made the news.

I kept testing new tools, too. And [music] in June, one of those tests felt very different. Fable 5 felt like unlocking superpowers. The way I describe it is this. Opus and the models I'd used before were an exoskeleton. Fable felt like Terminator. That's how big the difference felt in my work. There's even a commit in the repository called pre-Fable checkpoint dated June 10th. The next day, the history starts filling up with the 200 skill database.

This was the period when I expanded it to 200. And that brings the story back to Guild Wars. One of the design documents explicitly references its ritualist. The old idea of combining abilities was still driving the game. With Fable, the scale of what I felt able to build changed. Then, on June 12th, access was suspended. [music] US export controls had led Anthropic to suspend access to Fable for everyone. For me, it felt like mourning.

But the work it had helped establish was still there. The next commits describe continuing with Opus and Fable's implementation patterns. The workflow was literally called no Fable. On June 13th, the implementation log reached 200 out of 200. That still needed balancing and polish. But the project kept moving. Access returned on July 1st. When I tried it again, it felt so much weaker for my work that I stopped using it.

The game design took an unexpected turn, too. Somewhere along the way, I read a comparison between Guild Wars and Magic: The Gathering. I remember thinking, "Actually, yeah." Putting together a set of skills and figuring out how they work together did feel a lot like building a deck. So, I thought, "Why not try turning the skill tiles into cards?" But, with 200 skills, I was also facing a design problem. I didn't feel the game had the replayability I enjoyed in other Survivors games.

That led me to card backs. The idea was to introduce randomness into the cards and upgrades you got, giving you a different set of options to build around from one run to the next. You would still be making your own build, but the cards you received could push you to try a different combination. That shifted my thinking toward deck building, making something work with the options you have. The inspiration was still Guild Wars.

I had found another way to bring it into the game. Diablo III was another influence, especially its health globes. I liked how healing [music] out in the fight could make me more willing to dive into a horde and take a risk. I wanted that feeling in Code Mortis, too. Seeing a chance to recover health in the middle of danger, going for [music] it and keeping the fight moving. That was part of how I thought about the game's pacing.

Players helped shape those decisions, too. Our Crunch Fest Discord brought together players, people who saw potential in the game, and enthusiasts of AI and game development. Shout out to everyone there. I took constructive Steam criticism seriously and made changes in response to much chaos, unclear mechanics, runs that needed more variety. Bone Cathedral taught me a lesson. Its stalking boss was heavily inspired by Nemesis from Resident Evil 3.

I made him far too powerful. It took me about 8 hours to get through the level. Finally beating it felt amazing. Looking back, [music] I pushed the punishment too far toward rage quitting. The frustration outweighed the satisfaction I wanted from a stalking boss. I reduced the damage while keeping the encounter's mechanics. The change is there in the development history. My satisfaction after finally beating something was only one perspective.

I had to listen to what other players were experiencing. The first characters and the first images in general came out of a chat window. Getting them into a game was another problem. Backgrounds had to tile properly. Animation frames had to look like the same character, and I needed to process a lot of images. I couldn't get a model to do that consistently. So, I compromised. The wobbling you see in the early builds was my work around for proper animation.

Over time, I started using Pixel Lab's API. My workflow became GPT for concepts, then Pixel Lab to turn them into pixel art. Eventually, I got usable eight-direction sprites working. Later, having image generation built into Codex became a huge advantage in my workflow. Looking back after the move to an isometric look, I can admit it. The early game was ugly. I struggled with UI and UX, too. I learned a lot while making it.

Now, a concept could become a character I could actually move around in the game. Music mattered a lot to me. I spent a lot of time with music, and I wanted the soundtrack to carry the atmosphere of the game. Suno was indispensable for that. I could describe the emotion, the instrumentation, what I was hearing in my head, and work toward a track that captured it. This part is probably better heard than explained. >> [screaming] >> The soundtrack gave me a way to put that imagined atmosphere into the game alongside the visuals.

Back on the code side, that long cloud code period finally gave way to Soul 5.6. The change was practical, how far I could get with my tokens. With Claude, the limits had worn me down. I'd keep running into another message telling me I'd used up my allowance. Soul got the work done and I could keep going. For my workflow, it felt like I could work 24/7. Then came Astra. I'd already made several attempts at the game's interface.

With Astra, I rebuilt the entire interface in 2 days. That's why it's my first choice now. The final push toward release felt like a speedrun. But by then, something else had changed, too. How much of the software design I could delegate. Remember the corpse explosions and what I wrote in January about being the architect? With Fable, Soul, and especially Astra, the models began taking on more of the software architecture, too.

I still use my experience as a lead developer every day. Define the task, delegate it, review the result. I've become confident in that process. I'm still leading the work. The contractor has changed. This was a hobby project after hours. Looking at its scale, I don't know if I could have carried it through without AI. I'm not sure I would even have taken it on. AI took a lot of the mental load off me. The work felt lighter.

I had more room for creativity and bringing ideas to life. After the move to an isometric look, with the story coming together and 200 skills to build around, it started feeling like the game I'd wanted to make. I won't spoil the story here. I was also discovering it as a player. I'd get a mechanic design and the first implementation, then play it and find out what I liked. That helped me decide what to change next. I still think craft matters.

AI helped me learn, and a skilled artist using it will bring skills I don't have. And yes, I expect more AI slop. We've had asset flips, Unity slop, Unreal slop. To me, the problem is lazy execution. I already see it in AAA games, too. The tool alone doesn't decide the quality. Like a typewriter or a car, AI is a tool. I'm hopeful about what small ambitious developers can build with it. I expect the tools to keep getting better.

What we make is up to us. Now, I'm preparing Codex Mortis for version 1.0. Plan for September 24th. I tried all these models. This is what I made with them. Putting this retrospective [music] together, it hit me. My Kiro and Klein period was July 2025. Just over a year ago. Looking at what I can do now, it feels like going from an abacus to a rocket. Where will we be in another 18 months?