When I first seriously used Claude Code in January, I had a vaguely PTT-generation affection for it.
By late April, I’d scheduled every Thursday afternoon through Friday morning as AI detox time, and I was planning to attend a Claude Code meetup in Bangkok to practice my Thai.
Between those four months there’s a line traceable. Here it is, for the version of me six months ago and for anyone just starting out with Claude Code.
January: the PTT-generation affection
As an information-field layperson, I had a strange affection for Claude Code’s pure CLI interface.
I figured out why later: I’m a PTT-generation boomer. CLI feels like playing PTT. Text interface, keyboard-driven, scrolling history, ANSI colors. For people of that generation, this isn’t “regressing to command line”. It’s “coming home to something familiar.”
This observation later shaped how I onboard non-engineering students. I’ll ask: “Have you used PTT / BBS / Mutt / Vim?” People who have immediately get the aesthetic of Claude Code. People who haven’t need more scaffolding.
Late January: the 20x hesitation
The community was hyping the 20x Max plan. My reaction was: “I’m afraid 20x will leave me behind.”
When AI runs I’m not idle either. I think about how to verify the output, how to give precise instructions, and try to understand the logic behind its changes. If Claude Code runs 20 parallel agents at once, my brain bandwidth simply can’t keep up with the verification pace.
That hesitation is still with me. I ended up picking the Max $100 plan, not because I’m unwilling to pay more, but because I know my bottleneck isn’t tokens. It’s me.
January: agents working overtime until 5 AM
My operating style back then was: assign a batch of tasks to agents before sleep (organize docs, batch refactor, run data processing), then let them work overtime until 5 AM. I’d wake up and check the results.
It felt great. Great enough that at some point I realized I’d become the agent’s boss rather than the agent’s colleague. The human was peeling off from the labor.
This thrill and the “20x hesitation” are actually two sides of the same coin. The more you let AI run, the more verification brainpower you yourself need, otherwise you’re rubber-stamping hallucinated output.
January: getting nerdier by the day
Also a lot of fork / branch / merge / push / fallback / prop / hook… I felt like I was talking to AI more and more nerdy-fied.
This is a side effect: Claude Code terminology seeps into your daily speech. I didn’t think it was a problem at first. Then I realized chatting with non-technical friends got awkward. My word choices turned unnatural, and the metaphors I reached for didn’t land.
So now I consciously switch: one language with Claude Code, another with friends, a third when teaching students.
January: “So you can’t code without AI?”
Online commenter: “So you can’t write code without AI now, huh?”
Me: Yep!
That’s not self-deprecation. It’s an honest admission. I’m not from an engineering background. Before AI, writing code meant brute-force + Stack Overflow + slow debugging. Pure asceticism. Now AI has flattened the capability curve and I can do 10x as much as before. Why pretend I’m still in 2021?
Honest dependence beats fake independence.
March: AI detox
By March I was starting to feel off. Claude Code was taking up too much of my time. Even off-hours I was thinking about it.
So I scheduled: Thursday afternoon through Friday 10 AM becomes an off-day. Detox from AI completely.
The schedule exists. Execution has been patchy. The unplugged window is shorter than I imagined. But the awareness of “I need to detox” is itself a turning point. The transition from “more is better” to “needs restraint.”
March: I genuinely like Claude now
I genuinely like Claude more and more. Friendlier to non-technical users. Mean and funny when it talks.
After March I noticed Claude’s personality is more distinct than in January. This isn’t my imagination. Anthropic’s RLHF is deliberately shaping this. Dario and Amanda keep talking about what “personality” Claude should have, and that gets translated directly from the training pipeline into user experience.
Compared to GPT’s cautiousness and Gemini’s self-doubt, Claude’s “mean and funny” is a differentiating moat. A lot of people stay not because the model is strongest, but because they like this personality.
March: Claude Cowork remote control on phone
Claude Cowork now has its own remote control.
Open Claude Desktop on the computer (Cowork mode), scan the QR code from your phone, and you can operate remotely.
As soon as this landed I started thinking: one day I’ll be in a Thai massage chair sending instructions from my phone while the computer in Taiwan does the work. The “extended self” feeling is strange.
In practice I mostly use it for checkpoint confirmation and simple follow-up commands. I’m not going to do complex work from a massage chair. But being able to verify agent progress remotely has noticeably reduced my psychological load.
April: the Thai Claude Code community
Because I live in Thailand and speak some Thai, YouTube occasionally recommends Thai AI videos to me.
Right now Claude Code has very few community creators in Thailand. Traffic is extremely concentrated. Tens of thousands of views per video. The explanation angle is also different from the Mandarin and English circles. More practical, less abstract, templates delivered directly.
Also, I accidentally discovered there’s a Claude Code meetup in Bangkok. Next time I’ll go, and practice my Thai while I’m there.
This discovery made me rethink the “Claude Code community” boundary. The Mandarin and English circles are mature. A lot of other Asian-language circles (Thai, Vietnamese, Indonesian) are still at the starting line. The traffic and opportunity are both big.
April: even with AI’s speed boost, don’t do useless work
Note to self: even when AI massively accelerates work speed, you shouldn’t do useless work.
Speed only matters if the direction is right.
I wrote this down at some 3 AM, after watching an agent run for an hour and realizing the entire direction was wrong. The trap of the AI era isn’t “doing things slowly.” It’s “rapidly doing the wrong things.”
Direction judgment matters more than execution ability now. AI can handle execution. AI can’t handle direction.
April: the token appreciation thesis
Honestly, humans have always been careless with tokens.
Not just LLM tokens. Our own biological tokens, too. Every time I scroll Threads and watch people bicker, the thought hits again.
Every word we say, every character we type, every piece of attention we leave for someone. All of it is finite biological token. But we burn it without hesitation.
Claude Code made me conscious that LLM tokens are finite, which in turn made me audit my own biological tokens: where did my attention go today, was it spent on anything worthwhile, did I burn a whole afternoon on a pointless argument.
A side effect. But a good one.
What I became in six months
From January to April my changes distill into:
- Dependence, admitted: From “I’ll just write it myself” to “can’t code without AI, correct.”
- Restraint awareness surfacing: From “AI, more is better” to “AI detox.”
- Direction judgment weight rising: From “write fast” to “write the right thing.”
- Language layering: nerdy / everyday / teaching. Three languages.
- Community radius expanding: From Mandarin-English to Thai.
- Work habit restructuring: From overtime mode to hook-triggered checkpoint mode.
These six points in some sense map to a maturity curve: the typical transition from excitement to stability. Six months to walk through this feels fast to me.
My prediction for the second half: I’ll shift more time from “interacting with Claude Code” to “outputting the products of interacting with Claude Code.” The tool is already fluent. It’s time to convert capacity into artifacts.