I Watched AI Build My Website. Then I Worried About My Daughters' Math Tuition.

There’s a particular kind of quiet that settles in when you’re watching your kids do homework and you’re not sure whether what they’re doing still matters.

My daughters are seven. They’re working through their multiplication tables right now. Fours, fives, sixes. My wife drills them in the evenings, patient and methodical. It’s the same thing my parents did with me. It works. The tables stick.

But I’ve been sitting with this low hum of unease lately. Not because I think math is useless. I don’t. It’s more that I’m no longer sure the multiplication table is what the multiplication table used to be. And I’m not sure most parents, myself included until recently, have stopped to ask the question.

Something happened over the past few weeks that made it harder to avoid.

About a year ago, I tried building a personal website using Replit. For anyone unfamiliar, Replit is a browser-based coding environment where you can build and run projects without installing anything on your computer. It’s impressive for what it is. But I kept running into walls. Everything lived inside the platform. When I needed to connect external services, configure something at a deeper level, or install a specific package, I was stuck. I’d hit the ceiling of the sandbox and there was nothing to do but work around it.

I eventually gave up and shelved the project.

A few weeks ago I came back to it, this time using Claude Code inside VS Code. Claude Code runs from the terminal. That sounds like a small technical detail and it is, until you realise what it means in practice. It means the AI isn’t operating inside a walled garden. It can move around the actual system. Install packages, configure environments, connect to GitHub, run tests. It’s the difference between an assistant who can rearrange the furniture in one room and an assistant who has keys to the whole house.

What I built in a matter of days: a full personal website in two languages, English and Indonesian. An admin portal so I can edit and publish posts from my phone without touching code. A personal note-taking app powered by a separate AI model. Email subscription infrastructure. The whole thing deployed automatically every time I push an update.

Last year, none of that happened. This year, it did.

And I want to be precise about what I actually did. I wasn’t coding. I’d describe what I wanted, Claude would generate a plan, I’d bring that plan back to Claude to pressure-test it, then send the refined version to Claude Code to execute. Two AI instances effectively working in parallel, one for thinking, one for building. My job was to know what I wanted, describe it clearly enough to be actionable, catch when something came back wrong, and ask the right questions when I was stuck.

The skill that mattered wasn’t syntax. It wasn’t knowing how to write a function or configure a server. It was clarity. Knowing what you want precisely enough to direct something that can execute faster than you ever could.

Andrej Karpathy, who coined the term “vibe coding” a year ago, recently said the next phase is what he calls “agentic engineering.” The human role becomes orchestration: setting direction, reviewing output, making judgment calls. The agent handles execution. His framing clicked for me immediately, because it described exactly what I’d been doing without having a name for it.

I went back to watching my daughters with the multiplication tables that evening and the discomfort sharpened.

Here’s the thing about Singapore parenting, and I say this as someone operating fully within it. We are running a highly optimised program. Math enrichment from Primary 1. Chinese tuition because the curriculum is hard. Student care, then enrichment, then homework, then revision. The PSLE in Primary 6 sorts children into secondary school tracks. Parents know this, feel it, and respond rationally by preparing their kids for the sort criteria.

The program is coherent. It was designed for a world where execution was the scarce resource. If you could hold more knowledge, process it faster, apply it more reliably than the next person, you had an advantage. The exam system tests exactly that. So parents optimise for exactly that. It’s not irrational. It worked.

My own path ran through this kind of credentialing. MBA. Investment banking. A Master’s in IT with an AI specialisation, which I pursued deliberately because I could see where things were heading. Those credentials opened doors. I’m not going to pretend the system didn’t work for me, because it did.

But I keep coming back to a question I can’t dismiss. The credentials were proxies. Proxies for discipline, for problem-solving ability, for the capacity to handle complex information under pressure. Those underlying things still matter. The question is whether the same proxies still point to the same underlying things, or whether the map has drifted from the territory.

If AI handles an increasing share of execution, what is the exam actually selecting for now?

My wife and I haven’t signed the girls up for math enrichment. We talk about it periodically. There’s social pressure, the quiet kind, where you notice what the other parents at the school gate are doing and feel the gap. But we keep coming back to the same position. Not because we think math doesn’t matter. It does. But because we’re not convinced that more drilling of the same kind of content is the right use of the time and the money and, honestly, the stress.

What we do invest in: piano, public speaking, Chinese dance. I wrote about this in an earlier post, mostly through the lens of building confidence and character. But I’ve been thinking about it differently since the Claude Code experience.

Public speaking, at its core, is the practice of taking something that exists in your head and making it land clearly for someone else. That’s exactly what prompting is. Not the technical version with parameters and context windows, but the human version. Knowing what you mean precisely enough to make another party, human or AI, understand and act on it.

Piano is about holding a standard in your head and iterating toward it. You know what the piece is supposed to sound like. You listen to what’s actually coming out of your hands. You identify the gap and you work it. That loop, of having a clear internal model, evaluating output against it, and refining, is what I was doing the entire time I was building my website.

None of this shows up in a report card. I’m aware of that. I’m not fully opting out of the system. My daughters will sit the PSLE. They’ll need functional math and adequate Chinese. I’m not making a romantic argument that grades don’t matter.

I’m making a different argument. That the skills underneath the grades are shifting in relative importance, and if you’re only optimising for the grades without asking what the grades are actually selecting for, you might be preparing your kids for a world that’s being replaced faster than the curriculum can track.

We’re still doing the multiplication tables. We’re not doing math enrichment. We’re spending that time on piano and public speaking instead.

It might be wrong. We’ll find out. But at least it’s a deliberate call, not a default.

That feels like the right place to start.