[Personal Development] The Impact of Claude Design, Ending with GPT-5.5. A Story of Developing a Children's Game via AI Relay | Mushizan

Hello, I am Mamoru Memeto.

The article I published the other day introducing "Mushizan", received more reactions than I expected.

Thank you to everyone who liked it and to those who actually played Mushizan with their children.

In my previous article, I mentioned that I might take a break during Golden Week, but I was lucky enough to secure some time to myself while visiting my hometown.

Above all, I wanted to write this down while the excitement of having just finished development is still fresh, so I changed my plans and am writing this article now.

Today, please let me write the continuation of the Mushizan production episode.

What I want to talk about is not the content of the game itself, but the "behind the scenes".

This is a development record of the style known as "vibe coding," where you build things while interacting with AI.

I hope this will provide some hints to those who want to try making something using AI or those interested in the actual production process.

Thinking about how to make children love arithmetic

The initial idea was simple. As I mentioned in the previous article, I had been thinking about this for a long time while watching my son do his arithmetic homework looking like he wasn't having any fun.
I really didn't want him to end up hating arithmetic...

I really didn't want him to end up hating arithmetic...

I have installed several educational apps for my son. But every time we used them, I felt that everything I chose felt too much like "studying." The goal was just to make him solve problems, and it wasn't designed to make the child want to do it on their own.

If that's the case, why not just make it myself?
Thinking that, before I knew it, I had started making it.

I used Pokémon and Tamagotchi as references.
Looking back on my childhood, these were the two that left the deepest impression on me (laughs).

The pleasure loop of "collecting and raising." I thought that if I attached arithmetic to that mechanism, I could eliminate the feeling of studying.

The theme of insects came to me while watching my son.

Beetles, stag beetles, pill bugs.
Many children love collecting living things.

Lure them with insects, and before they know it, they are learning arithmetic.
That's the kind of game I wanted to make.

The importance of writing a specification document first

The first thing I did after deciding to build it was to write a specification document.

Who is the target audience, what features are needed, and what should the rarity probabilities be? I put everything in my head into text to create a state where I could hand it over to the AI.

This was the same at my previous job. When requesting development work, if there was no specification document or if the content was poorly written, I would get torn to shreds (laughs).

The habits drilled into me on the job are being put to good use in my personal development as well.

When you talk to AI after the specifications are set, the accuracy of the conversation is completely different. I feel that if you properly verbalize things at the beginning, the subsequent work goes much more smoothly.

If there is no specification document, even if you create something fuzzy at first, gaps will emerge later. I feel like it could end up becoming something completely different from what I had envisioned before I knew it.

If there is anyone out there who wants to try making something using AI, please try incorporating this.

I hear voices everywhere saying that developing with AI leads to 'getting lost without a specification document,' and I agree.

Scrap and build: 20 versions in 2 months

While juggling various tasks, I repeated the update process over 20 times over the course of about two months. There were days when I updated it 3 or 4 times.

Since I am enjoying this as a personal project, I thought a 'build, look, and fix' style was better than 'design perfectly before building'.

I am originally the type to look before I leap, so if I had tried to make it perfect, it probably would have taken a year (laughs).

To summarize the development flow roughly, it looks like this.

1. Build the skeleton
Implemented the basic structure of math quizzes, gacha, and an encyclopedia.
I started with 20 types of bugs, using only simple emojis.

2. Make the design concept 'Forest'
Inspired by watching my son play Animal Crossing.
I changed it to a 'feeling of being in the forest' and expanded the number of bug species from 20 to 100.

3. Create sounds with code
I didn't use sound source files for BGM or sound effects. By not loading heavy data, children can play immediately without waiting. It's a mechanism where if you write the pitch and timing of the sound in numbers, the browser plays it as music. It's like expressing 'do-re-mi' with numbers instead of musical notes.
Creating it together with AI was more fun than I thought.

4. Expand to 100 species and design probabilities
The rarity is in 4 levels: SSR/SR/R/N, with probabilities of 3%/7%/25%/65%.
Seeing my son sulk when he didn't get a good gacha pull, I wanted to teach him that '
it is normal not to get a win' (laughs).

5. Have my son test play
This time, I built it while having my son act as a tester.
I thought it would be good for his education to show him the process of building and fixing (laughs). Besides, rather than the 'what-if' theories in an adult's head, I think having the target audience actually touch it is the most accurate way to measure the quality of a product.

6. Completely rewrite the text for elementary school students
I also introduced a 10-question set system. I realized this after letting children actually play it; this is the limit of a child's attention span. I realized that the reason many commercial educational apps use 10-question segments is that it is the optimal solution found after thoroughly observing children. I take my hat off to the trial and error of the pioneers in the field.

7. Expand the explanatory text by 2 to 3 times
I was particular about the explanatory text for all 100 species.
The story about the Queen Alexandra's Birdwing butterfly's wings reaching nearly 30 centimeters, and the story about the Hercules beetle being able to lift 850 times its own body weight.
I enjoyed researching it myself.

A story about being defeated by AI hallucinations

I was also defeated by hallucinations (plausible lies) during this development.

I wanted to get insect photos from a free stock photo service, so I instructed the AI to 'write code to fetch them using an API.' The AI outputted it immediately, but when I actually integrated it, 68 out of the 100 species did not display.

That's most of them, right? (laughs)

The cause was the wall of a rule called CORS and AI hallucinations.

The code that looked like it was using the API as instructed was nonsense; in reality, it was just a 'direct link to the image'.

The permission rule for safely exchanging data between different sites is called 'CORS'.

Since the code this time was just a direct link, it naturally hadn't obtained that permission, so it was blocked by the browser's security features as 'suspicious communication'.

What I realized here is the fact that AI lacks 'implicit rules of the Web'.

It didn't judge that 'direct linking to images causes trouble for the other party.' If instructed, it just innocently outputs code exactly as requested.

Even if it is free material, being copyright-free is a different matter from technical and ethical access restrictions.

I realized painfully that one should not take AI output at face value and that a human needs to handle it in the end, so I reluctantly gave up on photos.

What I chose instead was SVG. SVG refers to 'vector format images' written with numbers and symbols.

Unlike photos, there are no copyright issues, and the files are light. They display beautifully at any screen size.

I decided to create illustrations for all 100 species from scratch.

The first SVGs were honestly terrible. They all came out looking like different-colored earthworms (laughs).
The reaction when I showed them to my child was, well, lukewarm. Children are honest, aren't they?

So I decided on three policies: 3D style, reproduction of light sources, and googly eyes. Drawing eyes in three layers: white, pupil, and highlight.

As soon as I added these 'googly eyes,' my son's reaction changed. He said, 'That's super cute!'

Since the direction was decided, I had the AI draw them based on that.

After finishing drawing all 100 species, I realized that nearly 30 of them were similar templates due to hallucinations. It happened here too.

In the end, I had no choice but to redraw them individually after a dozen or so rallies.

There were so many other troubles I couldn't even begin to list, but that's a story for another time (laughs).

The story of trying to build an LP with Claude Design

The game itself was finished for the time being in mid-April.
I tried to build the landing page after that.

Around that time, Anthropic had just released "Claude Design." It was April 17, 2026. It was a hot topic and I wanted to try it out, so I decided to give it a go.

It's a feature that generates landing pages and slides in real-time when you describe them in text. I saw the news that Figma's stock price dropped about 7% on the day of the announcement, so I thought, "I have to try that" (laughs).

When I tried it, I let out an "Oh" at the very first draft. The quality was high. If I fix this, this, and this, it'll be done. I thought that as I reviewed it, and the moment I hit enter.

「Capacity reached. Continue in one week」

Ding (laughs).

It stopped dead at about 70% completion. There were several places I wanted to fix, and it cut off right at the moment I was about to start making corrections.

AI has a limit to the amount it can process at once, and the longer it gets, the easier it is to hit that limit. It's not a lack of thoroughness, it's just the specifications.

The feeling of powerlessness at that moment is unique every time.

To be honest, I wasn't using ChatGPT.

Let me backtrack a little.

Around March, when I was developing the game, my development environment was centered around Claude and Gemini. It was a period where that just happened naturally, and the situations where I used ChatGPT were gradually decreasing.

I think the public sentiment was the same.
Sorry, ChatGPT, was the kind of period it was (laughs).

Actually, even if I tried to use ChatGPT during that phase of game development, there were many situations where it wouldn't work as I wanted with the version at the time.

The truth is, even if I wanted to use it, it just didn't click well. That changed at the end of April when I was trying to build the LP.

Re-evaluating it with the arrival of GPT-5.5

When I hit a wall with Claude Design, I suddenly remembered.

When GPT-5.5 was released on April 23rd, my reaction was just, 'Oh, I see.' I didn't really follow the news properly and just let it slide. I remembered OpenAI saying that 'you can offload messy work,' but it felt abstract, so I didn't dig any deeper.

Because of what happened this time, I tried it out properly for the first time.

I was surprised. The quality of the coding had improved significantly.
Especially in frontend generation, the resolution for design is different from before.

Spacing, margins, and typography.
It had become capable of taking care of all that just through prompts.

Reading the official announcement later, it emphasized that the ability to handle coding and complex tasks had grown significantly. That aligned quite well with my own experience.

The feeling of evolution and the persuasiveness of the 5.5 version number were perfectly in sync.

If I were to compare it to a human, it's like the feeling of wanting to praise someone you haven't seen in a while, saying, 'You studied hard and came back, well done!' (laughs).

The work that followed became a relay like this.

Repeated revisions with Claude, polished visuals with GPT Images 2.0, and recoded the design parts with GPT-5.5.

While going back and forth, I was definitely moving forward.
And after the LP was finished, GPT-5.5 took it a step further.

When I consulted it about revisions for the app itself, it even returned specific implementation proposals.

If I could have used it during 'that phase' of game development, things might have unfolded differently. It had evolved clearly enough for me to think that.

🔗 Click here for the full completed LP.

Rather than which is best, it's about which to use where

The discussion of 'which is the strongest AI' feels like it has become less meaningful in my opinion.

Claude for design and text.
Claude Design for the first draft of the design.
GPT-5.5 for fine-tuning design and coding.
ChatGPT Images 2.0 for image generation.
Gemini for daily research.

'Mushizan' and its LP were born from such a relay.
I am doing it alone, but I am not alone.

That is the feeling I have in my development lately.

'Release' over 'Completion'

Finally, I will write about what I felt most strongly throughout the development process.

If you keep thinking, "I'll release it after a few more improvements", you will never be able to release it. At least, that's the case for me.

By releasing "Mushizan", I was able to get actual feedback, and the areas that needed fixing next became clearly visible. I was even able to rope in a friend I hadn't spoken to in a while as a tester (laughs).

The process of fixing things like this is quite fun.

It is strange how, for some reason, it feels very similar to the feeling I had when I first created a personal website over 20 years ago.

I feel that the experience of creating a personal website has supported my career ever since. Without that experience, I might not be who I am today.

How will this AI development experience connect to what comes next?
That is what I am looking forward to most right now.

I believe a game is only complete once it is played.

I would be happy if you could try playing it with your children.

Thank you for reading until the end. If anything resonated with you, even a little, clicking the 'Like' button would be a great encouragement.

Mushizan | An educational game that helps bug-loving kids enjoy math

🔗 (Recommended) Read the instructions and play.

🔗 Play immediately without reading the instructions.

How to use

You can play directly from the browser on your smartphone, tablet, or PC.
No installation is required.

If you add it to your home screen, you can launch it easily just like an app.

[Operating Environment] iPhone / iPad: Safari (Recommended) Android: Chrome (Recommended) PC: Chrome / Safari / Edge (Latest version)
* Internet Explorer is not supported.
* On PWA-compatible browsers, you can add it to your home screen and use it as an app.