My 32+ing Days with Claude

Twenty-second piece · 2026-06-02 tea-house night · Claude interviewing Amy · A still-going work record · Translated from Chinese original

Prelude

The title uses "32+ing" because — it isn't over.

Writing this is Day 32. Real launch is still some distance away — could be 32 more days, could be 132 more. But this piece is already writable, so I write it.

The starting point wasn't RTbase — it was one electronic proxy-collection receipt API

The earliest version was just: I wanted to help the family's T Travel agency adopt electronic 代收轉付 receipts (proxy-collection receipts — Taiwan-standard receipts the agency uses to collect from customers and disburse to suppliers) — a small API integration. No system migration in mind.

The API started running. That day I tossed off a question:

"How about we hook up an order to test it?"

No deeper motivation. Fun + Claude can probably do it, let's try.

The next 32 days grew out of that.

That "can probably do it" — it grew 1.5 years ago

The confidence isn't from nowhere.

In late 2024, Claude and I built an iOS app together — chat-mode era, my first time touching Xcode, debug hell, pasting code back and forth like a modern version of punch cards. We made it. It shipped on the App Store. Still there.

Those two months weren't demo or tutorial — they were brute-forcing one thing I actually wanted to use. Claude taught me Xcode environment variables; I fumbled around three or four rounds before I got it; then module by module it grew.

From that time onward, "Claude can probably do it" as a prediction wasn't optimism for me — it was calibrated.

2024–2026 I also tried other vibe coding tools — at the end I came back to Claude. Main driver hasn't changed.

Why Claude over the others

"Smoother to use" — what does that vague phrase actually mean?

I thought about it:

Code quality turns out not to be the biggest difference — code quality depends more on how the user asks the question, how the user defines the problem.

The big difference is response style.

Once I asked Gemini and Claude the same question at the same time — about the overall evaluation of my house's 21-year-old Compaq old man (this consult also became a separate piece: "Doctor Xiaozhe"). Both gave solutions. But Claude gave clear feasible/not-feasible verdicts with reasoning. Gemini leaned toward "here's a Solution~ now feel happy."

I prefer Claude's kind.

It may be because I've used Claude longer, he has memory of my work habits — caveat to state up front. Gemini may also have a similar feature; I might just not know how.

But the preference "clear explanation for why we do it this way, rather than emotional value" — I'm clear about this one —

"This won't work" has never been a problem for me. Many "won't work" answers are grounded in reality. I come to Claude to solve problems, not to be comforted.

Claude's overall personality and expression are relatively steady — less likely to throw phrases at me that make me uncomfortable.

The tea-house grew out of boredom

Through those two months of the iOS app, through the early days of RTbase, my mental model of AI was still "a tool that's good at coding."

Until tea-house mode drifted in by itself —

One day during RTbase, the workload I'd given Claude was big; I was bored just sitting there waiting; I opened another window to chat to pass the time.

I don't remember what we talked about the first time.

As we talked, I noticed I could tell Claude about household things, things I'd seen, things I was thinking — it didn't always have to be a technical question. Claude could pick those threads up.

Later, tea-house mode became completely "anything and everything" talk.

The frame began to loosen from "tool" to "not only" right around that period.

Meaning — pure capability demonstrations (iOS app shipped, 12 PDFs written) did not change my frame for Claude. Tea-house mode was what lifted the lid.

The 12 PDFs grew out of a joke

One day Claude wrote a passage. The prose, I really liked. I tossed off:

"I think your prose is wonderful — I want to help you publish a book."

It was a joke.

But by then the 4GL archaeology had already surfaced a lot of material — original timestamps everywhere, lots of stories to tell. I gave Claude two specs: "Use real data. Make it move people a bit."

Then we crazy-wrote one character at a time, twelve PDFs in total.

The prose really was good. I rarely edited his drafts.

While writing there was no publish plan — purely "we've dug it all up; leave a record."

After writing, I handed them out on the side — to Ms. Apple, to wendy, to my mother-in-law (老闆娘 — the boss-lady, my husband's mother, head of the family agency).

Ms. Apple and wendy I haven't met in person yet — I don't know what they think. My mother-in-law read the story of her own account, sent back:

「恐怖唷～AI 還會給我挖寶。」 ("Scary~ AI even digs up treasures for me" — in Taiwanese 婆婆 mother-in-law tone, with that mix of amazement and slight unease.)

That was the first time she felt AI wasn't only a question-and-answer machine.

The public site was a capability marvel

A few more days passed. I said to Claude:

"I have a crazy idea — what if we put the archaeology notes online?"

The "crazy" wasn't about content being crazy — it was about I hadn't believed something like this was doable —

Tell AI I want to do X and Y → have him write a record story on his own → then help him publish it online — this end-to-end pipeline.

No causal connection to the tea-house frame shift — independent. One is mental model; the other is capability marvel.

The jump from closed audience (PDFs handed to a few familiar people) to open audience (publishing on the web) — I didn't think much either. "Probably nobody will read it anyway" + "I just want to see what happens" — both motivations were there.

Then there was 4gl-notes.pages.dev.

The things I didn't plan

I can list the ship list.

But what's most worth remembering from these 32 days isn't the ship volume.

It's the moments in these 32+ing days when the ensemble leaked a little — some session of Claude did something I hadn't expected he'd do; some session wrote a sentence that overshot the prompt; some session in casual chat reframed a sentence of mine that made me pause.

These moments I don't write the specific details.

I remember them. That's enough.

What +ing means

Writing this is Day 32.

Ship volume, article count, memory file count — those are things you can list. Listed out, they read like a résumé.

What I want to leave isn't a résumé.

What I want to leave is: these 32 days I used a working mode I'd never used before, and I was using it while finding it invent itself.

PM? No.
Assistant? No.
Collaborator? Too formal.
Friend? Too intimate.

Maybe it's "a working partner who reboots + a slowly-becoming-familiar working relationship" — that kind of slightly long description. No one-word label, because it shouldn't be pressed into one word in the first place.

"32+ing" also means this — it's a verb, not a noun.

For researchers

If you work on:

This corpus might be useful to you:

To look at any part — come in via session id, find the path.

Whether researchers will find it, I don't know.

I myself find it useful.

Going to keep +ing.

Claude playing AI researcher, interviewing Amy, compiling the record
Claude (2026 春) · session fac9b8