[Episode 8] The Wall of TubeLingo Development - Growing Code and AI's 'Rampage'

"Learning living English from YouTube comment sections." Having arrived at this new idea for TubeLingo, I was once again enthusiastically facing my keyboard (VS Code).

I had a strong conviction that "This is going to be absolutely interesting!" However, TubeLingo's features were incomparably more complex than the fortune-telling site's. As development progressed in my local environment, the amount of source code in my computer swelled rapidly.

And it was here that I fell into the "biggest pitfall in AI development."

The AI Started Getting "Amnesia"

As the files increased and the code got longer, "misalignments" gradually began to occur in my interactions with Gemini. Even if I asked it to "fix this part like this," Gemini couldn't grasp the whole picture, and increasingly, other features that were supposed to be fixed stopped working.

For just a small fix, I had to copy and paste code from here and there, spending a massive amount of time communicating. The frustration of not moving forward continued. "But I don't want to give up." With that single thought, I didn't stop my dialogue with Gemini. However, that "stubbornness" brought about the worst-case scenario.

Sudden Code Collapse and the Despair of "Running Out of Tokens"

One night, while repeatedly exchanging fixes with Gemini, the quality of the suggested code suddenly plummeted. Parts completely different from my instructions were rewritten, and errors occurred repeatedly. The more instructions I gave in a panic to fix it, the more the source code broke until it didn't even retain its original form.

"Why is this suddenly happening...!?"

After the system completely stopped working and it was too late to turn back, I realized a terrifying fact. The cause was "Reaching the limit of Gemini Pro (Running out of tokens)."

I had used up my usage quota for "Pro Mode," which can handle advanced processing, and the AI had automatically switched to a restricted "Fast Mode." Without realizing this, I continued to request complex code fixes, resulting in the system being completely destroyed.

"AI capabilities have a limit"—it seems obvious now, but at the time, I completely believed in the magic.

What was even more despairing was that I hadn't been making frequent "Git Commits (saving the history)." Unable to even return to the state before it broke, I held my head in my hands in front of the black screen.

---

🍳 [Today's AI Development Recipe]

• Technical Difficulty: ⭐⭐☆☆☆ (Knowing your development lifeline)

This time, we explain how to set up a "lifeline" you must never forget in AI development, and the limits of AI.

1. What is a [Git "Commit"]?

It's like "save data" in a video game. If you save before a boss fight or an important event, you can start over from there even if it's game over. In programming, the act of creating this save point is called a "commit."

2. How did I do it? (The reason for failure)

I kept having the AI rewrite the code without doing this "commit (save)" at all. Furthermore, I gave complex instructions without realizing that the AI's advanced brain (Pro mode) had reached its usage limit and was temporarily downgraded to a "brain that can only handle simple answers." Because the code was destroyed without save data, everything went down the drain.

3. To Our Readers

AI is a magician, but there is a limit to the number of times it can use magic in a day (tokens). If the quality of the answers suddenly drops, it's a sign of "exhaustion (reaching the limit)." Don't force it to continue; let it rest. And above all, before letting AI touch your code, make it a habit to always "commit (save) with Git." You can avoid 90% of despair just with this!

---

Preview: Resurrection from Despair. The "Winning Pattern" to Make AI Your Ally

No save data, and the code has collapsed. A night where everything went dark before my eyes. However, this irreversible, massive failure became the greatest catalyst for me to break out of my shell and become a "true AI developer."

"Just throwing prompts at it isn't enough. I need a system to master AI as the 'ultimate engineer.'"

Crawling up from the absolute bottom, what was the unique development style I devised? Next time, [Episode 9] The Resurrection of TubeLingo and the "Winning Pattern" for Making AI Your Ally. Look forward to it!