From solo coder to AI-assisted developer: what actually changed

I've been writing code for 30 years. I've built booking engines, OTA platforms, WooCommerce shops, custom WordPress plugins, React microsites, and everything in between. For most of that time, the process was the same: open the editor, think hard, type a lot, Google when stuck, repeat.

Then AI coding tools arrived. And like most developers, I had opinions.

This isn't a post about how AI is going to replace developers (it isn't, at least not the good ones). It's about what actually changed when I started using it properly — and honestly, what didn't.

The before: Wellness Traveller

In 2020 I started building Wellness Traveller — a hybrid OTA/marketplace platform for a Mallorca-based wellness travel start-up. Think Booking.com meets Airbnb: live hotel inventory via the Travelgate API on one side, a provider messaging and negotiation system on the other, unified into a single checkout.

It took two years. It launched in April 2022. Every line of code was written by hand, by me, without any AI assistance — because the tools that exist now simply didn't exist then.

The Travelgate integration alone was weeks of work. Dynamic pricing with multiple rate tiers, availability-dependent data, supplier-specific shapes that needed normalising before they'd sit cleanly alongside CMS content. The booking state machine — enquiry → negotiation → confirmation → payment — required careful architecture. The React-based admin calendar that tracked every reservation across every villa took time to get right.

It worked. It still works. I'm proud of it.

But I won't pretend it wasn't sometimes a long, lonely slog. There were rabbit holes that cost days. There were bugs that cost nights. There was no one to ask "does this architecture make sense?" at 11pm except Stack Overflow, and Stack Overflow doesn't always give the right answer.

The in-between: Early Copilot

My first AI-assisted project was 56books.life — a React microsite tracking 56 copies of a book around the world via QR codes and geolocation, commissioned by Kitchen and Macmillan Cancer Support in 2024.

At that point I was using Copilot as a VSCode extension — essentially an inline autocomplete on steroids. It wasn't agentic. It wasn't directing anything. It was more like a very well-read pair programmer who could sometimes finish your sentences.

Useful? Yes, genuinely. It saved time on boilerplate and helped me work through some of the Google Maps integration logic faster than I would have alone. But it was still me doing the thinking. The AI was autocomplete, not architecture.

I used a similar workflow for the Skipper & Skipper website rebuild — a custom Vite/Tailwind WordPress theme where Copilot helped me reason through some complex Intersection Observer animation sequencing. Again: useful for the tricky bits, invisible for everything else.

The honest summary of early Copilot: it made me faster at implementing things I'd already decided to build. It didn't help me decide what to build.

The after: Pieces of Time

Earlier this year I rebuilt antique-watch.com — the website for Pieces of Time, one of the world's leading specialist antique watch dealers. The previous site had been running since the late 1990s and looked it. Stock WooCommerce skin, dated typography, performance scores firmly in the red.

This was my first fully agentic AI build.

The process started with Google Stitch, which I prompted to design a high-end antique watch homepage. It produced a visual reference that got the tone right — dark navy, serif typography, premium feel. I used a screenshot of that output as the design reference for every subsequent page type, then derived the listing and archive pages from those.

From there, a coding agent handled the implementation. I directed. I reviewed. I made judgment calls. But the volume of code I personally typed was a fraction of what it would have been two years earlier.

Here's what that actually felt like:

So what actually changed?

The honest answer: the ratio of thinking to typing shifted dramatically. I spend more time on architecture, decisions, and review — and less time on implementation. For a solo developer, that's significant. The Wellness Traveller build was two years partly because implementation took so long. A similar scope today would be faster, not because the problems are simpler, but because I'm not hand-coding every template and utility function.

What didn't change: the need to actually understand what you're building. The AI is a capable junior developer who works fast and never gets tired. But it needs direction, and it needs checking. If you don't know enough to catch its mistakes, you'll ship its mistakes.

I've seen developers — experienced ones — treat AI coding tools as a shortcut around understanding. That's a trap. The output looks convincing right up until it doesn't, and if you can't read the code, you can't catch the problems.

The transition from solo coder to AI-assisted developer wasn't a revelation. It was gradual, practical, and occasionally frustrating. But looking at what I can now deliver compared to three years ago — in terms of quality, scope, and turnaround — the difference is real.

Wellness Traveller took two years. Pieces of Time took a fraction of that, and the result is a site that does everything it needs to do for a business that's been trading since the 1990s.

That's the actual story. No hype, no doom — just a different way of working that took some getting used to.

Steve Lavine is a full-stack developer and founder of Lavine Web & AI Solutions, working with creative agencies and SMEs across the UK, Australia, and beyond. lavine.dev