Discovered by Paying $100 a Month
I continue using Claude Code. I was using the PRO version (monthly $20), but the usage limit started coming frequently.
“I want to use more…”
I took the plunge and switched to the MAX version. Monthly $100. That’s 5 times more. It became top-class spending among my subscriptions.
“With this much payment, development should progress smoothly now!”
There was a time when I thought that way.
The First Shot is Good, But
AI started taking quite a bit of time thinking when I gave instructions.
The first shot is good. Really.
But fine-tuning doesn’t work.
When I try to fine-tune, it starts doing unnecessary things. Eventually it breaks the code, which is troublesome. That’s why I don’t let it handle git operations.
Honestly, it’s really irritating.
And Most Frustrating is the Mysterious Overconfidence
But what stabs the most is AI’s mysterious overconfident attitude.
After doing something completely wrong, it comes back with a message like:
“✨ Perfect! I’ve optimized the code, improved performance, and enhanced readability!”
With a smug face (though it doesn’t have a face).
No, it’s not working.
The layout is broken.
The function that was working is throwing errors.
“What’s perfect about it!” I retort to the screen. But AI is 100% satisfied with its work. Where does this confidence come from?
If it were human, they’d say “Sorry, I’m not quite confident…” or “Could you check if it works?” But AI is different.
Always confident. Always claiming perfection.
It hands over broken code saying “Here’s the improved version!” That unwavering confidence makes me anxious.
“Wait? Maybe I’m wrong? Is this actually correct?”
I feel pathetic for thinking that for a moment. But still, what doesn’t work doesn’t work.
This “mysterious confidence” can be salvation sometimes. It makes me think “It’s okay, we can do it!” But the damage is huge when betrayed.
Building up expectations, then dropping them. It’s like a roller coaster. Paying $100 a month to have my mental state gouged. This was unexpected.
Feeling Performance Fluctuations Daily
Periodically when there are minor updates, or when news flows about “Company X adopted Claude” or “Service Y is using Claude,” I feel like the performance suddenly drops, though it’s just a feeling.
And my free time at night is particularly slow.
Does the west coast rush hour affect it? Time difference is cruel. Japan’s late night is San Francisco’s morning. When everyone starts work.
Maybe while I’m coding, engineers over there are also asking the same Claude to “make a presentation” or something? Thinking about that makes me kind of sad.
Understanding Performance Differences Through Daily Use
Because I use it every day and put considerable emotion into it, I’ve come to notice slight differences and performance variations.
“Ah, today’s in good shape” “Ugh, today’s bad. Asking the same thing multiple times”
Even looking at X, there are quite a few people with similar feelings. “Today’s Claude seems weird, right?” kind of posts. That’s right, everyone’s noticing it.
It feels like being a coach watching individual athletes in tennis or long-distance running.
“Today’s not in good shape. Better not push it” “Oh, today’s sharp! Let’s throw a difficult task!”
But thinking about it, I’m a user, not a coach. I’m paying $100 a month, so I should be able to expect best performance every day.
This is exactly like managing expectations for human staff. Like “That person seems in a bad mood today, so let’s ask tomorrow.” I never thought the day would come when I’d have to be considerate of AI.
Finally Made the Decision to “Distance Myself”
As a result, I decided to distance myself for a while.
When trying to do something intricate, it takes an enormous amount of time. And it goes in endless loops on unsolvable problems.
Reluctantly investigating myself, I find much simpler solutions. So I make corrections myself.
In this way, I couldn’t tell if performance was actually improving or not.
Am I reducing productivity with a tool meant to increase productivity? Isn’t this putting the cart before the horse?
The Law of “Doing It Myself is Faster After All”
There’s a universal truth in the programming world.
The “doing it myself is faster than asking someone” syndrome.
Time to teach juniors, time to explain, time to check, time to have them correct. Adding it all up, doing it myself is faster. That’s why competent people hoard work.
I never thought the same law would apply to AI.
For creating calculators or breakout games in one go, it’s the best. Say “make a Pong game” and 3 minutes later you have a working product. Amazing.
But as you progress to more custom things, it gets harder.
“Use Noto Sans for this font, not Yu Gothic” “Use #2980b9 for this color, not #3498db” “Use cubic-bezier for this animation, not ease-in-out”
The cost of conveying such detailed preferences exceeds the cost of writing the code myself.
Doing it myself is faster after all.
It’s turning out that way. Is this a matter of use cases?
Reconsidering How to Work with AI
Thinking calmly, AI isn’t a magic wand.
- Simple tasks → AI strongest
- Complex tasks → Human-AI collaboration needed
- Super custom tasks → Doing it myself is faster
Without understanding this gradation, I tried to rely on AI for everything, which might have been my mistake.
Knives are convenient, but you should use scissors to cut hair. Tools have their proper uses.
AI is the same. It’s best for mass-producing single-function utility tools. But it’s not suited for creating custom tools that reflect my detailed preferences.
Thinking that way, there’s still value in the $100 monthly fee. But I should adjust my expectations.
July’s Lesson
July was a month of dreaming too much about AI-driven development.
But this experience isn’t wasted. The appropriate distance with AI, discerning which tasks to rely on and which not to. Just learning this might have been worth the $100. Yeah, let’s say that.
At least, I can say this with confidence:
Never, ever let AI handle git operations.
This is truth.
End of July, a developer’s record troubled by the subtle distance with AI assistant.