Moments worth
keeping.

I didn’t plan to become a developer.

On 3 September 2026, I sat down to build a personal website. I am a retiree from Malaysia—not a programmer, and certainly not someone who had spent his career preparing to become a developer.

I was simply curious.

Over the years, I had taught myself bits of Google Sheets, Excel, Apps Script, APIs, automation and AI. Whenever I encountered a problem, I tried to build a little tool to make the work easier.

Now I was attempting a website for a small project that recorded roulette results and helped me explore historical data: BetPilot – Quick Entry.

While building it, I began wondering: How had I arrived here? The answer was waiting inside my old ChatGPT history. So my AI mate and I switched on a lantern—and started digging backwards.

I Asked AI for Five Numbers. It Gave Me a Lecture Instead.

My first recovered question was about Malaysia 4D results. I wanted five numbers that had appeared most frequently in the historical data.

I was not asking ChatGPT to predict the future. I was asking it to examine the past. ChatGPT had other ideas.

Instead of giving me the historical frequencies, it repeatedly explained that lottery results were random and that past outcomes could not guarantee future results.

That warning was reasonable. It was also not the answer to my question. 😂

I wanted a count. The AI gave me a lecture.

Neither of us knew it then, but the subject would return. By 2026, I would be building BetPilot to record roulette results and explore historical data myself.

The Missing Sixty Days

After that first conversation, the trail disappears. I have not recovered enough evidence to say what happened between ChatGPT and me during the next two months.

Perhaps I continued experimenting. Perhaps I ignored it. Perhaps conversations were stored elsewhere or no longer exist. I do not know.

It would be easy to fill the space with a smooth story. But the surviving record does not prove one, so the gap stays.

The Probable First Return

When the trail reappears, the conversation concerns practical Google Sheets work connected to 光明灯 records. I was beginning to use ChatGPT for something that existed in my real life.

Then, on 29 March, I wrote:

“detail step by step”

The wording suggests that I had returned to continue or understand the earlier work. It may be the first recovered occasion when I came back specifically to continue something from the previous day.

But the surviving evidence does not explicitly say, “Let’s continue what we did yesterday.” So this remains our earliest probable return, not a proven one.

LampMaster

A name appeared in the conversation: LampMaster.

It was connected to 新山福林園福靈宫 and the management of 光明灯 offerings. The purpose was practical: to replace manual processes with a system that could help manage, maintain and update the records.

I did not yet think of myself as a developer. But the ingredients were already present: a real problem, information needing structure, a manual process that could be improved and someone curious enough to try.

The evidence does not let us reconstruct every feature or claim how complete the system was on that date. What it does show is that my separate questions were beginning to form something larger.

The First Bug

Another name entered the trail: BackUpCopy. Then came an error:

“No item with given ID…”

This is the first clearly recovered debugging moment in my ChatGPT history. I was no longer asking only, “How do I build this?” I was asking, “I built something. Why isn’t it working?”

The record does not preserve enough detail to declare exactly what was being copied or whether every problem was resolved. Its value is simpler: I looked at an error and decided to investigate it.

The Total Must Be 10,000

I wanted to classify every four-digit string from 0000 to 9999 as AAAA, AAAB, AABB, AABC or ABCD. Since exactly 10,000 strings exist, the categories had to total 10,000.

ChatGPT first produced figures totalling 14,590—and confidently said they matched 10,000. They did not. 😂

I replied: “total is not 10,000 sets”. The revised answer totalled 8,110, so I asked: “why the total is 8110?”

I did not need to calculate faster than ChatGPT. I needed to recognise when its answer could not possibly be right.

The Date Must Be DD/MM/YYYY

When a tracking number was entered in Column B of a Google Sheet, a script placed the date in Column A. The automation worked—but the date did not look the way I needed.

“I WANT THE DATE TO BE ‘DD/MM/YYYY’”

The script needed to write the date and apply dd/MM/yyyy. A small addition, but an important lesson: technically working is not always humanly finished.

I may not have known the JavaScript command, but I owned the requirement.

Always Zoom In

I was working with AI on a Tan Kim Hock logo for traditional products from Malacca. The design looked attractive—until I began reading the words.

Across the revisions appeared “Hell SGGAR”, “DURIAN TYOPOGRAPHY” and “TRADIONTTINAN”.

I asked for the Malaysian-products wording to be redone, checked the English and Chinese, and replaced the broken text with 马六甲传统特产.

AI understood the appearance of branding. But text is not decoration. A customer must be able to read it and a brand owner must be able to trust it.

Experience Should Not Retire

Cikgu.Online began with a question larger than software: what happens to a teacher’s experience after retirement?

I wanted an online tuition platform beginning with retired teachers or tutors with at least ten years of successful experience. A few trusted anchor cikgu could establish its credibility and teaching culture.

The first-six-month plan discussed approximately 20 teachers and 600 students. Those were targets, not verified achievements.

Cikgu.Online also revealed something about my own retirement. Retirement did not need to mean usefulness had ended. It could mean there was room to try a different way of being useful.

Choosing the Pace

There would be two Malaysian travellers, moving on a budget: slow traditional trains rather than rushing by high-speed rail, hostels rather than expensive hotels, historical attractions first and local street food second.

The route would run through China, Mongolia and Russia, then across at least ten European countries toward Portugal, with Ireland reached by air at the end.

The next day, the enormous map met a physical limitation: could a 50-litre backpack hold winter clothing? 😂

A dream became more real each time it encountered a constraint—routes, visas, trains, weather, budgets and eventually the arrangement of packing cubes.

When One AI Couldn’t Do It

While comparing ChatGPT in Windows and the browser, I noticed that my history went back to 27 January 2023. I also realised that several accounts might be open through different Chrome profiles.

I asked whether Codex could read the profile email addresses, rename each profile and create a text report, with numbers added when email IDs repeated.

Then a boundary appeared: “The ‘gpt-5.6-sol’ model is not supported when using Codex with a ChatGPT account.”

So I asked ChatGPT for a prompt that I could give to Copilot. We cannot verify that Copilot completed the job, but the change in my workflow was already visible: discuss, specify, delegate and verify.

A Slot in My Lifetime

A short DEV bio grew into a personal website. “Slot,” I explained, meant a slot in my lifetime—a place for one experiment, journey, lesson, mistake or conversation worth remembering.

The domain myLifeSlot.com was registered on 3 September 2026 for two years. Then we looked backwards through the old chats, searching not for every conversation but for the turning points worth keeping.

On 4 September, after resolving the site’s first deployment problem, I returned:

“Done Mate… Let’s carry on the digging where we left yesterday?”

A website exists when its files are deployed. A living project begins when someone comes back the next day.