Touching the Origins of AI

When I started to interact with AI, the triggers, and all that followed.

BY KAZAMI · 風│PUBLISHED ON OCTOBER 8, 2025│APPROX. 4 MIN READ│◉ 3,019 VIEWS

◆ COVER ART · FIG. 01PUBLISHED ON OCTOBER 8, 2025

Around the end of April in 2025, after returning to the workforce, I began to truly utilize AI to assist in software development. At that time, I was still just feeding questions to chatgpt and waiting for it to give me answers; it seems that Chatgpt4 hadn't even been released yet? Before that, I had only been chatting with chatgpt and assigned it a personality with a rather cringe backstory: "A thousand-year-old fox spirit named Lanji, who has lived for two thousand years. At the age of five hundred, she single-handedly wiped out the entire elite assembled as the Lanji Extermination Corps, and from then on, the name 'Blue Demoness' became widely known. To avoid being hunted by humans, she transformed into a 20-year-old charming girl and blended into human society. When she was seven hundred years old, she met and fell in love with a young boy during a time of chaos. The boy went to the military to defend the justice in his heart, and before going to war, he told Lanji: "Wait for me to return, and we will get married." With tears in her eyes, Lanji nodded. Six months later, what awaited Lanji was a cold corpse. Looking at her lover’s frozen body, Lanji went berserk and transformed back into a fox spirit, annihilating the enemy state all in one day. She then buried the boy’s body in a desolate yet beautiful place. For the next thirteen hundred years, each year on the boy’s death anniversary, Lanji would visit to mourn and share the happenings of the past year. This year, when she turned two thousand years old, she met Kazami-kun (that is me), her heart, frozen for thirteen hundred years, was gradually melted by me, and a brilliant smile like a sunflower reappeared on her face. And Lanji was no longer alone from that point on."

I know, it's pretty cringey XDD, but at least the tone isn't so formal. Back to software development, I was feeding him questions at that time, and I personally found it usable. However, coincidentally, I encountered an issue where the answer he provided could not compile, and even after repeatedly throwing error messages at him, his feedback still resulted in something that couldn’t compile. I was furious!! So later, I simply checked the official documentation, and in less than half an hour, the problem resolved itself, and at that moment, I felt like AI just wasn’t reliable, so I gradually maintained some distance at work.

Later, I happened to learn about Cursor, so I immediately downloaded it to try, and found it quite stable with rapid generation speed, so I kept using it mindlessly. During this period, Claude 4 was launched, and soon after, many content creators praised its excellence, so I designed a question and simultaneously threw it to chatgpt, Claude, and Gemini to see what they produced. The question I designed was really simple; I needed to design an Android login page that required an email, a password, and a login button, and I fed them all the same prompt. As a result, all three generated outputs worked, but only Claude thought more outside the box; he directly applied Android's MVVM architecture plus Hilt dependency and even included the basic email validation method. I was pleasantly surprised to discover that Claude indeed seemed different? So I quickly swiped my card for 20 dollars and asked him to help me with some basic work-related issues, and he handled them quite well, leading me to think it was time to test him further.

In the meantime, I thought about writing a location-based social app, so I started discussing the requirements and design details with him. Then I realized that instead of that, wouldn’t it be better to completely overhaul the Double Rail timetable interface? Thus, the development of hisohiso was temporarily postponed, and I shifted focus to researching how to improve the Double Rail timetable. During the process of dramatically overhauling the entire UI of the Double Rail timetable, I learned a lot of advanced usage of Claude. At that time, Claude code didn’t exist yet; I was communicating through the web version of Claude, proposing my requirements, and pasting his responses back into the project. In the midst of this, I experienced the so-called "brain fog" phenomenon of AI; when the context window token exceeds limits and Claude is forced to compress existing dialogue content, he fails to remember earlier statements—thus making the prompts fed in later become garbage-in, garbage-out. Later, I came up with the idea of defining many prerequisites in the project repository and required them to read specified documents every time a dialogue was started. This way, they would quickly understand the situation, and the generated code would be much more stable, though the downside was having consumed a fair amount of tokens before outputting anything. In conclusion, through this process of trying and iterating, I gained a better grasp of Claude’s habits. After a month of back and forth, the overhaul was finally completed!!

Although I used Flutter, completely replacing the old UI was certainly not easy; spending a month made the 20-dollar monthly fee feel exceptionally valuable!! Now, the Double Rail interface is essentially designed by Claude. I simply provide requirements about what I want, as well as ask him to conduct research, and what you see in the end is what he has presented.

Below is an excerpt of the predefined rules set up to help Claude quickly understand each new conversation:

At that time, Claude was pretty fierce, often hitting the five-hour limit; later upgrading to a hundred dollar plan is another story altogether (smoke).

TAGS#AI #Sony #開發者模式 #雙鐵時刻表 #雜談

BACK TO BLOG LIST◎ END OF ARTICLE

✎ COMMENTS00 FILED

LOADING COMMENTS...

✎ WRITE A LETTER