Episode 1. A Person Who Learned to Code for Five Minutes Tries Building an AI App
Here is the honest version of my coding experience. Once, I followed some addition and subtraction lesson in Gemini's learning feature for about five minutes.
Then I thought, "Why am I even doing this?" and closed it.
That was it.
If you asked me to properly explain what a variable or a function is, even now I would only have a vague sense of it. I still could not explain it precisely.
And yet somehow, I am now releasing an app. This story starts there.
What Annoyed Me Every Day
At the time, I used a lot of AIs. Each had its strengths and weaknesses, so I kept Claude, GPT, and Gemini open and used all three.
I used Gemini because it came with 2 TB of Google Drive storage (then one day it suddenly became 5 TB), and Claude was my main tool because it was smart.
But even with Claude Pro, there was always pressure from the usage limit. The Opus allowance disappeared so quickly that I could barely use it, so I mostly used Sonnet. (I still felt it was the smartest.)
I was holding out on GPT's free plan. It was just a reference tool. Whenever I asked GPT something, the answer was so long and drawn out that I simply kept using the free version.
To be honest, I did not really understand the differences between the models at the time either.
One AI would say something, and I would go to another and ask, "This one says this. What do you think?"
I would copy Claude and Gemini's answers into GPT, then copy GPT and Gemini's answers into Claude.
Paste it here, paste it there. It sounds trivial, but repeating it over and over was incredibly annoying.
I also tend to organize everything as a "project," with instructions and related source material. The problem was that I had to create the same project separately in every AI.
It was the same work, but I had to make a project in Claude, another in GPT, and another in Gemini.
Every time I turned on my computer, I also had to open each AI, find the project again, and then find the right chat.
(I later solved part of this with Chrome tab groups. If I grouped tabs by task, the pages I had been using would reopen the next time. But that only solved the problem of "reopening the windows." All the other inconveniences remained.)
Google Drive was another problem. All my material was in Drive, but I could not add Drive files directly to a project's source material.
So every time, I downloaded a file from Drive and attached it again.
GPT supports Drive integration now, but it did not then. I do not think Grok did either. And GPT still makes you enter a file link instead of letting you drag and drop.
Even adding Google Drive files as attachments was inconvenient.
A Door I Opened by Chance — Google AI Studio
<Image 1-1. Google AI Studio>
I do not remember exactly how I found it. One day, I happened to open something called Google AI Studio.
It told me to "describe the app you want."
Oh, so this is that thing. Something like OpenClaw that had been getting attention—apparently you could tell it what to do in natural language and it would do it.
Was that why people were suddenly looking for Mac minis?
I heard the term "vibe coding" a little later on YouTube.
The idea came to me immediately. Why not build an app that combined the GPT, Gemini, and Claude windows I kept open every day?
Solving just the things that bothered me most would be enough.
- No need to create the same project separately in every AI
- Add Google Drive files directly to project sources
- Include a notes feature
- And this was a little ambitious: let the AIs exchange answers and produce a summary
That was the first picture I had in mind.
Starting with Two Five-Dollar Payments
Apparently I needed APIs. To put the AIs into an app, I had to get a key—an API key—from each company.
I worked through it one step at a time by asking Claude or Gemini.
Gemini billed afterward, and while experimenting inside AI Studio, I could even use it for free.
So I paid only OpenAI and Anthropic five dollars each. I had bought keys for GPT and Claude.
The First Connection, and the First Strange Moment
I told it to create three slots on the screen. Now it was time to connect the three AIs.
When I sent the same question, GPT, Claude, and Gemini would each answer on a single screen.
Like the real apps, I wanted to label Claude models such as Opus, Sonnet, and Haiku as high, default, and fast.
<Image 1-2. Models by slot>
So I wanted the exact model names and asked each AI. That was where I hit the first obstacle.
They gave me names that were obviously several generations out of date. They were genuinely confused about which model they themselves were.
I will get into that properly in the next episode. "AI does not even know itself" was the first sign of how many times it would leave me baffled later.
(Continued in Episode 2)