“With this, I could make endless games!?” — I’ve thought the same thing three times in the past three years. And all three times, I was let down.
It’s been two years since I left my last job. In that time, I’ve been consuming indie games, envisioning my next game, and above all, researching the latest AI. In reality, I started my research after my previous game shipped in November 2022, when I suddenly had a lot more free time. And right at that exact moment, Stable Diffusion and ChatGPT were released. Being there in real time for a turning point in IT history was huge.
With ChatGPT, I tried game research and bouncing ideas off it for planning. The output varied wildly depending on the prompt, and back then the model only had data up to around 2021, which made things pretty tough. In a highly specialized field like mine, it couldn’t produce stable output, but for general exchanges it was usable enough. More than anything, the fact that something I’d previously had to do either alone or with other people could now be done anytime, 24/7 — that was significant.
With Stable Diffusion, I was blown away that I could generate unlimited images locally. “With this, I could make endless games!?” — I thought. But as I kept using it, it could only produce the same kinds of images. It depended entirely on the model data, the prompt precision was rough, and even after generating thousands of images, not a single one was usable — a sad result.
In April 2025, a major ChatGPT update brought huge improvements to image generation. “With this, I could make endless games!?” — I thought again. But once more, my enthusiasm wilted the more I used it.
The accuracy of photorealistic images has improved, sure, but generating images usable as game assets is abysmal — probably because there’s too little training data in that domain. Illustration-style output has extremely low color temperature, and the patterns seem limited. With the public backlash against generative AI acting as a deterrent, I don’t see rapid progress anytime soon.
In October 2025, Gemini’s image generation got a major update. The ability to maintain consistency was a big deal. Remove this person from the image, turn them to face right — being able to treat the source image as a base was excellent. Before, it would generate similar images but change all sorts of other small details. That problem was solved. I thought, “Now I can use this for games!” But unfortunately, the quality still isn’t there for direct use as game assets.
I went in with high hopes, but in the end, image generation wasn’t something I could use.
What We Decided Not to Use
At Impachi, we don’t create art or sound with generative AI. This is a firm decision. When it comes to art, I’ve been crafting it with great care throughout my career, and current AI doesn’t meet even my minimum standards. And of course, given the intensity of public backlash, the risk is just too high.
I didn’t mention it earlier, but the same goes for sound. I spent some time with Suno — the current frontrunner — and actually made some tracks, but it’s hard to use for games either. Songs with vocals are listenable enough even if the backing audio quality is poor. But game music is instrumental, and the low audio quality is unbearable. Most composition AIs barely respond to prompt instructions at all, so you can’t change the melody, can’t change the chords, can’t change the timbre. Sound is arguably what I value most in games. It’s improving, but it won’t be usable for years, and I have no intention of using it.
What We Do Use
Let me document how I’m handling development and company operations as of April 2026.
Right now, I’ve settled on Antigravity + the Claude Code for VS Code extension. I’m combining a development environment with an editor and Claude, which excels at Japanese input and output. By the way, the game currently in development runs on Unity. I’ve set things up so Claude can operate Unity, but that’s getting into the weeds — I’ll save it for another time.
As for what I have it do — first and foremost, coding. Today’s AI excels at generating and reading text, and since programming languages are text data, the compatibility is excellent. There’s also a wealth of training data from publicly available code. In the coding domain specifically, I believe Claude is the undisputed leader for professional use as of 2026.
I was an IT engineer up until 15 years ago, so I’m not incapable of writing code, but game-specific coding is a high bar. Still, by teaming up with Claude, it’s become easy for me to come up with ideas and have Claude implement them. “So you don’t need engineers?” — people tend to think that, but it doesn’t work that way. Our CTO Furumoto has far more technical knowledge, and having an engineer around helps a lot. No matter how capable AI is, it can’t drive itself. It won’t come up with interesting specs or designs we haven’t asked for, and it won’t implement them either.
Next, I also use it for organizing game planning and specifications. At first, I used the web version, bouncing ideas off it and compiling things, but it would often change specs on its own or misread things. Brainstorming game design means adding and subtracting ideas while gauging the other side’s reactions — being able to do that without another person is incredibly valuable. On the other hand, I was still manually extracting text from the chat and compiling it into text files. With the integrated development environment, that happens automatically. It also maintains consistency with other files and can check them. If information is missing, it can research the web too. How great is that? In games, you need to create things that don’t exist in the world yet. In that case, AI can’t produce the right answer from zero. It can output whether something is “fun” or “not fun” as statistics derived from text humans have written in the past, but I can say definitively that it lacks the ability to create something unprecedented with conviction that it’s fun. Of course, it’s excellent at reasoning from written text, so for someone like me whose strength is zero-to-one ideation, it’s a great partner.
Beyond that, it’s also useful for running the company. Starting with paperwork for incorporation, invoice generation, various applications, payroll, accounting, task management — and it even generates this website. For the website, I tried having it auto-generate content from a large volume of my personal history and thinking that I’d fed it, but the result was so off-putting that I write everything myself. I also built an automated English translation pipeline. You could use separate services for each of these tasks, but having everything consolidated is a huge deal for a newly founded company with limited funds.
Making a Game with Three People
Impachi started with three people. Without AI, we’d probably need more members. That said, can three people with AI make a game at the scale we used to build with dozens? Absolutely not. We don’t use AI for art or sound, so we need outside collaborators. And even for design and programming, AI won’t produce anything beyond what we instruct.
What has changed significantly is that design ideas and features we used to painfully cut because the implementation workload was too large are now becoming feasible. Most recently, we’ve been experimenting with having AI playtest and write up its impressions. We’d never release without human testing, but it seems useful for statistically analyzing balance and tuning enemy behavior. We can also expect it to compress the schedule for time-intensive processes.
I think of AI as a tool that quickly delivers 70-point work. In games — entertainment — 70 points is fatally low. The remaining 30 points come from original game design, world-building and narrative that move people emotionally, and the quality of art and sound.
There are risks in publicly stating that we use AI, but I’m putting it in writing deliberately. While borrowing AI’s strength, I hope we can make games that are creative, high-quality, and genuinely fun.