January 2025: Discovering the frontend and vibe coding

I’d never built a website until the start of the year. Through building Marty, I had to get into it (for the personal site, I didn’t want to use a no-code solution like WordPress, out of a tech-person’s ego).

Claude Sonnet 3.5 had just come out. I thought:

Go on, I’ll build myself a personal site. With the right tech choice and AI it’ll be fast.

A few prompts, some generated code, a deployment on Vercel, and there it was. A working site in a few days.

The problem? I understood nothing of what I was looking at. The SEO was poor (I didn’t even know what it was), the performance mediocre, the design generic. But it worked and it was better than what I thought I could do, so I was happy.

AI had given me the illusion of having built something.

Experiencing Amara’s Law

Roy Amara formulated this now-famous observation:

We tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run.

In January, I was right at the peak of the curve. AI = magic solution to everything. I recommend this video by Xavier Delmas which explains it very well.

Mid-2025: The “serious” redesign

After several months of frontend work on Marty, I realized I could do better.

Second version, still with AI, but this time with more experience. I was able to ask for the right things, structure properly, optimize certain points.

The result? Better, but still not great. And above all: hours spent generating, debugging, adjusting code, and paying the associated tokens.

All while remaining dependent on a tool.

September 2025: The template

While working on improving Marty’s artisan profiles, I came across astro.build/themes.

In a few minutes I realized I’d spent days (even weeks for Marty) coding things that were already available online, better made and free.

I found a template that did exactly what I wanted: SEO-optimized, top performance, accessible, clean design, understandable code.

Half a day later, my site was live. Better than my two previous versions.

What I took from this

AI is a rocket: it’s incredible how much it can accelerate things when you know exactly what you want.

But if you don’t know exactly what you want, it’s like going very fast in all directions and ending up in the right place by accident.

AI as an amplifier

Over the past few months I’ve consistently noticed:

On subjects I know well (backend, containerization, data, mlops): Time hugely saved with AI

On unclear subjects (advanced frontend, …): Time lost going in circles

AI amplifies what you give it: clarity or confusion.

In the end, I should have asked myself a basic question back in January:

What do I actually want this site to do?

Send a quality, modern signal, and be able to share content easily.

Very different from coding a site from scratch with AI. That serves no purpose and worse, it consumes time and money.

The optimal solution already existed, for free, built by people who master the subject: a simple template.

AI doesn’t replace thinking

With Claude 4.5 just released and even more capable, the temptation to just let it run (it’s genuinely remarkable) can feel real.

But you need to find a balance. Going forward I’ll ask myself these questions:

  • Does this solution already exist somewhere?
  • Do I really need to code this?
  • What’s the real value I want to create?

In short, think for 5 minutes before rushing headfirst at the latest technology.

TL;DR: AI is a powerful tool. But it doesn’t replace thinking. Sometimes the best solution isn’t AI-assisted development, it’s simply thinking for 5 minutes.