This article was originally written in Japanese and translated into English with AI assistance. Please note that some expressions may carry nuances from the original Japanese.
🇯🇵 日本語版はこちら / Japanese version
At the end of the previous article, I wrote: “I started design at the end of December 2025 — four months ago — and the first product still isn’t done.”
This time, I’ll walk through what happened during those four months, in chronological order. Over these four months, I barely made progress on my main business — the seven AI systems. Instead, three products I had never planned got off the ground. I say “unplanned,” but all three are serious products in their own right. I need to walk through it step by step just to make sense of it myself — so I decided to write it down.
January — The First Detour Began with Data Migration
At the end of the year, I purchased a new MacBook Pro to move toward AI system development. Migrating data from the old iMac was the first hurdle. But this turned out to be hell beyond anything I had imagined.
My iMac, used for many years, had accumulated an enormous number of files, never properly organized. Files with identical names and identical sizes were somehow scattered across different folders. I couldn’t tell which were the originals and which I could delete. I had probably backed things up to an external HDD at some point, because the same files existed on both the internal and external drives. I tried Rsync, I computed file hashes for comparison, I tested various methods while consulting with Claude — but the amount of manual work was simply too much.
And at a certain moment, I thought: “If there were a tool that could semi-automate this migration work, it would not only move my iMac cleanup forward, but also have value for other people with the same problem.”
Just like that, the data migration work itself stopped, and development of a tool to support data migration began. This is what would later become the first product, called DataMigrator. I began the migration so I could finally move toward my main business — yet somehow I ended up building a tool for the migration itself. My main business slipped even further out of reach.
End of January — A Business Trip Tore Me Away from AI
Just as the DataMigrator concept was beginning to take shape, a business trip for my main job came up. For over two weeks from late January to early February, my development work came to a complete halt.
Anyone who is doing personal development as a side project probably understands: continuing development during a business trip for your main job is simply not realistic. You can’t do personal development on a company laptop, and carrying two laptops (company and personal) on a trip is too heavy, both physically and mentally. As a result, business trip periods forcibly tear you away from AI.
I think this is a common pattern for side-project developers. The pace at which I produced 20,000 lines of design documents in two weeks at the end of the year — why did it suddenly stop? The answer is simple: I was physically cut off from the environment where I could focus and work alongside AI. You can do text-based brainstorming, but you can’t do real work. On top of that, I had enthusiastically subscribed to a high-tier AI plan intending to “go all-in on AI-assisted development,” but during the trip I couldn’t use my tokens. I spent the trip almost in tears.
February — The Second Unplanned Venture Began from Garbled Text
After returning home, I tried to resume DataMigrator development. But here, I encountered another problem. Files I had saved from past Claude conversations were garbled and unreadable.
When I looked into it, the cause was differences in character encoding. UTF-8, Shift-JIS, CP932 — there are many types of character encodings, and unless you identify which encoding a file was saved in, you can’t read it correctly. I hadn’t expected such pitfalls to be hiding in everyday Japanese text files that I had been using without any awareness.
And then, the same thing happened again. “If I’m struggling with this, then someone else must be too. Let me build a tool that detects character encoding and converts files.” Just like that, DataMigrator development stopped, and development of a second product — utf8conv — began.
My main business drifted even further away.
March — Tax Filing Consumed Several Whole Days
Just as utf8conv was nearly taking shape, mid-March arrived. Tax filing season.
I registered as a sole proprietor under Soul Resonant Works on March 23, 2025. In other words, this was my first tax filing. On top of that, it was the more complex “blue return” format. Even with accounting software, at first I could barely understand what the software was asking me or what the displayed terms meant.
“What is an employment income deduction?” “How are the numbers on this withholding statement from my employer calculated?” “What happens if I file a blue return?” “What is loss and profit offsetting?” — a general search would turn up plenty of explanations on the internet, but to truly internalize them in the context of my own situation, I needed to go one step deeper.
This is where AI shone. Every time an unfamiliar term appeared on the accounting software screen, I asked Claude. From the general meaning to the specific application to my case, I worked through it conversationally, one concept at a time. It was probably faster — and more insightful — than reading an entire book.
Over several days, I somehow got through the tax filing. Once again, several working days for my main business evaporated.
End of March — The Third Unplanned Venture Dropped in at a Dinner
On March 27, having finished tax filing and finally ready to return to my main business, I went out to dinner with an acquaintance.
The food at that restaurant was so delicious that, at the register, I said to the owner: “My name is Kanazawa, and I carry around what you might call the ‘Kanazawa Guide’ — a collection of my recommended restaurants, like a private Michelin Guide. This food was so good I’d like to register this place as a starred entry.” The owner replied: “Please, by all means, make it three stars.”
At that moment, the words “Kanazawa Guide” — which I had just spoken aloud myself — started running through my head from that day on. “Wait, could I actually build this?”
The next day, I consulted Claude. That was March 28. From there, I had Claude as my sounding board every day, refined the concept, put together a business plan, and embarked on implementation. This became the third product, called “maiguru.” Instead of getting back to my main business, I found myself launching yet another product.
The products I had been considering were primarily on-premises software products, designed to run without an internet connection. But maiguru is software implemented on the internet. It’s in a completely different direction from anything before, but this one too is moving toward realization through collaboration with AI. In less than two weeks from that March 27 dinner, development has progressed to the point where a proof-of-concept prototype is running on a production-environment server.
Over Four Unplanned Months, I Had Built Up Basic Capacity
Laying out the timeline makes it clear: over these four months, implementation of the seven AI systems — my main business — has barely progressed. Completing the requirements definition for M1, the first of the systems, is where I currently stand. Compared to my original projections, I am clearly behind.
On the other hand, the three products I had never planned — DataMigrator, utf8conv, and maiguru — are each moving. DataMigrator has progressed from requirements definition to design, utf8conv is nearly complete, and maiguru has started running in production.
The level of completion varies across the three, but since I have been working with AI constantly, the fundamentals of vibe coding — better ways to write prompts, better ways to move forward, better ways to ask questions — have been converging. Each time I took on a new product, I became more efficient than the last. This has also meant personal growth.
The Pitfall of an Era When AI Can Build Anything
Why did things end up this way? Let me try to think it through.
In a word, I believe AI has caused such an explosive expansion in Capability — the sheer range of what’s now possible — that ideas can now be turned into reality almost instantly. For someone like me — who tends to move through each day following whatever ideas happen to surface — an era where everything shines brightly has arrived. You could say products now take concrete shape almost immediately.
Let me touch briefly on my own programming experience. In my student days, I had a chance to learn C. The result: I dropped out at Hello C World (a program that displays “Hello C World” on screen when run — the very first thing beginners are made to program in C). In that first class, when told “#include <stdio.h> — just think of it as a magic spell,” I wasn’t the kind of person who could accept things as magic spells. But I also didn’t have the energy to research and understand it myself. When told “See? It’s complaining about a Syntax error,” my state was: what on earth is Syntax in the first place? When told “This pointer here…” I thought: isn’t a pointer the arrow you move with a mouse? My grades were terrible.
Fundamentally, I am terribly unsuited to logical thinking. I can’t draw paths like “if you do this to this, this happens” or “if you compute this after this, you get that.” And I didn’t understand the grammar either. I was completely useless at programming.
And yet, someone like me can describe what I want to build, discuss it with AI, and have the AI handle architecture and coding. Honestly, even if I restarted my life three times and devoted all three lives entirely to software development, I could not reach the volume of architecture and coding that has been produced in these four months.
Not long ago, even if I thought “a data migration tool would be useful,” I had no implementation skills, so I would simply have given up. Same with “if only there were a character encoding detection tool.” Same with “if only there were a Kanazawa Guide SNS.” There was a high wall between ideas and implementation.
That wall has been lowered by AI. Or rather, to be precise — for someone like me, the wall has essentially disappeared. It’s been lowered so much that I can build whatever I think of, one after another. As a result, I can no longer focus on my original main business. I believe this is a problem that many people developing with AI will face going forward.
Fortunately, in my case, the “three products that emerged as byproducts” ended up serving as the foundational skills and capacity for moving forward on the seven AI systems. But that’s only with the benefit of hindsight — it wasn’t what I was aiming for from the start.
From the next article, I was planning to cover each of the three products born over these four months, in order. However, as I began the blog, I realized there is no English version of the homepage. The “unable to reach main business” pattern activates again. Next time, as a special edition, I’ll bring you the record of building a multilingual site in collaboration with AI.
About Soul Resonant Works
Soul Resonant Works is a solo venture developing seven local AI systems.
Starting from zero programming experience, the development is progressing through collaboration with AI.
🌐 Soul Resonant Works:
→ https://www.sr-works.net/en/
📝 This blog publishes the entire development process as a serialized journal.
If you found this article useful, please share it.
Leave a Reply