AI Development is Managing a Genius Alien — The Reality Those "Tetris in 5 Seconds" TikToks Don't Tell You

Table of Contents

AI Development is Awesome! (Or So I Heard)

“I asked AI to ‘Make Tetris’, and it created a working game in one shot!”

I’m sure many of you have seen such magical videos on TikTok and dove headfirst into the rabbit hole of AI development, thinking, “Coding is going to be a breeze from now on! I’m going to be a 10x engineer starting today!” I was one of them. But when you actually try it yourself, reality hits you hard. “Huh? It doesn’t go that smoothly…”

Here’s the thing: Tetris works well because its specifications have been strictly defined historically and are universally known. However, in the actual software development trenches we face, that’s not the case. Specifications are fuzzy, requirements shift like quicksand, and often we can’t even articulate “what we want to build” in the first place.

Suddenly, are you familiar with the sitcom “The Big Bang Theory”? It features geeks who have abnormally high IQs but zero social skills. Among them, the character Sheldon Cooper stands out. He is undoubtedly a genius, but an extremely specific, difficult person who doesn’t understand common sense, has intense obsessions, and struggles to communicate properly with normal humans. Sarcasm flies right over his head, and he has a dedicated “spot” on the couch that nobody else can touch.

Current AI is exactly “Sheldon”. He won’t work unless you explain everything from scratch, with excruciating detail. However, only when you manage to provide “perfect instructions” that satisfy him, he hammers out overwhelming productivity that humans can never imitate. Recently, I’ve lost track of who is managing whom, so I end up asking the AI, “How should I instruct you?” …Wait a minute. Am I the one being used here?

The “Trick” Behind the Magic

I have been developing using AI for about a year now, and one thing constantly crosses my mind: “The amount of context you have to spoon-feed AI is more than you imagine.”

• Your definition of “normal” is miles apart from an AI’s definition. Expressions like “make it look good”, “balanced”, or “within a reasonable range” are equivalent to telling the AI to “roll the dice”. You almost never get what you pictured.
• Existing processes, requirement definitions, constraints, how to proceed with tasks… you need to explain everything repeatedly with the patience of a kindergarten teacher. It’s truly at the level of “Listen, this is how you hold scissors. No running in the hallway.”

Once you get the hang of the AI’s quirks, it does become easier, but the ultimate bottleneck is always “tacit knowledge” and “verbalizing your own thoughts”. This is exactly the same challenge as assigning tasks to humans, but it’s amplified because your counterpart is an inflexible machine. Well, eventually, AI that can “read the room” will appear through multi-modal learning… but until then, we’re stuck hanging out with a “genius alien who can’t take a hint”.

My Epic Fails

When I first started AI development, I tried throwing tasks that weren’t fully baked in my head, thinking, “just figure it out”. Usually, the AI starts running wild in a direction I never predicted. I’d watch it go and sigh, “No, that’s not it at all.” I’d repeat this cycle, only to realize hours later that I’d spent the entire afternoon staring into the abyss, wondering, “What am I even doing with my life?” I did this multiple times. It’s a ridiculous situation where I spent more time wrestling with prompts than I would have spent just writing the code myself.

In the end, “Tasks that are too vague to assign to a junior engineer cannot be assigned to AI either”. This is the golden rule. Thanks to my job, I was somewhat used to defining tasks clearly, but I suspect many people drop out here and conclude “AI is useless”.

I also tried the trendy “Spec Driven Development”. This is a method where you interact with the AI to create requirement definitions, system designs, and task lists first. I thought “This is the holy grail!”, but when I actually tried it, plans that looked solid at a glance had massive holes, or I ran into unexpected blockers during execution. Once you hit that wall, you end up right back in the “running wild” state I mentioned earlier.

And the ultimate temptation is: “It doesn’t matter if I don’t understand the code as long as it works”. Proceeding without checking the implementation because “Hey, it runs!” is like saying “I don’t need to know which wire triggers the bomb”. If you don’t grasp at least what is implemented in which file, the code becomes unmaintainable garbage the moment something breaks. I made this mistake once when building a tool for myself, and eventually had to rewrite the whole thing from scratch. A mountain of code that works but you’re too terrified to touch—that is the black magic created by AI.

How I Use It: Micromanagement up to “One Step Beyond My Comfort Zone”

After trying various workflows, my current optimal solution is: “Use AI to boost what I can already do myself”.

1. Check and Expand My Domain: Do I understand exactly how to implement this at the code or infrastructure level? If not, I first use AI to study and expand my own knowledge. Before letting him write code, make him my tutor.
2. Design the Process Myself: Don’t be lazy. Face the problem and decide “I would proceed in these steps”, creating a detailed execution plan. If you leave the process design to the AI, it builds something that looks like it works, but is often fundamentally different from the ideal architecture. You are the architect; the AI is just the carpenter.
3. Micromanagement: At the execution stage of individual tasks, I specify the How (tech stack, coding style, tests to implement, files to change, etc.) and strictly control the AI’s behavior. Since typing out long explanations takes time, I start with rough instructions, and if the output is bad, I provide more detail, iterating until it’s right.

I use him to boost my output within the range I can understand. This exquisite distance is the key to transforming AI from a “difficult-to-handle genius alien” into the “strongest partner”. However, this also means the leverage is applied to your base intelligence, so I sometimes feel like the game balance of software engineering is starting to break.

Conclusion: Entering the “Infinite Loop”

Software development is changing dramatically due to AI. Especially for my use case, “working up to about one step beyond my comfort zone” has become hyper-efficient. In this timeline, the era where a single “superman” can build products at explosive speed—like the author of “Everything Claude Code” recently discussed on GitHub—is probably right around the corner.

In my case, tasks that used to take 2 weeks can now be finished in 1 day. So, what do I do with the extra 13 days? I use AI to learn new technologies again. To use AI more efficiently, I learn about AI, and have AI create tools for that…

I wonder if I am already trapped in an infinite loop of “using AI to use AI”.

Well, it’s fun, so I guess it’s fine! See you later.