I spent fifteen years learning to build websites the slow way. Design first, then WordPress, Elementor, a stack of plugins, a hosting panel, a cache to tune, a security surface to babysit. I was good at it. Roughly half the client sites at the agency I freelance for came off my keyboard. So when I say I went AI-first, understand it is not because the old way stopped working. It is because the gap between an idea and a shipped, paid-for thing got small enough to walk across in an afternoon, and I did not want to pretend that had not happened.
What fifteen years actually taught me
Designer first. Then WordPress and Elementor, because clients wanted to edit their own pages. Then agency work with Made by Proxy, where I picked up hosting, security and maintenance for other people's clients. Then real estate marketing with Crealys: landing pages, virtual tours, CGI, brochures and branding for property developers who needed to sell units before the concrete set. Every one of those jobs taught me the same lesson from a different angle. The hard part of this work was never the typing. It was judgement. Knowing what to build, what to cut, what a page is for, why a headline fails, where a checkout loses people. Fifteen years bought me that judgement. AI did not.
That distinction matters, because the loudest voices on both sides get it wrong. One camp says AI writes the code now, so craft is over. The other says AI output is slop, so nothing has changed. Both are selling something. What actually happened is narrower and more useful: the cost of turning a decision into working software fell through the floor, and the value of making good decisions went up.
The moment it clicked
I did not believe the hype until I shipped something real with it. Not a demo. A product with users and a Stripe account. The first was Orbit, my own agency operations platform: a form-based CMS for static sites, a leads inbox with AI-drafted replies, a Google reviews tab, an analytics view, an ecommerce module. Solo. It runs the very site you are reading this on. Building that ten years ago would have meant a team or a year. I did it in weeks, in the evenings, because AI collapsed the distance between "I know exactly what this needs to do" and "it does it."
Once I trusted the loop, the products stacked up. TrackTrendy, a privacy-first analytics platform, where I eventually replaced a third-party engine with my own and passed 1218 of 1218 validation checks at cutover. SendMerlin, an email marketing tool with a real sending pipeline: a worker that drains 500 recipients, ramp and auto-pause, one-click unsubscribe to the RFC 8058 spec, over 250 tests between unit and end-to-end. Roundmark, a design review app where you pin comments on a live site. LeadMerlin, which watches Reddit and Hacker News for buying signals. Eight products live now, and I am one person.
AI did not make me a better developer. It made the distance between deciding and shipping short enough that judgement became the whole job.
Where AI helps and where it absolutely does not
Let me be specific, because vague enthusiasm is useless. AI is very good at the parts of a build that are laborious but well understood: scaffolding a component, wiring an API route, writing the tests I would have written anyway, converting a design into clean markup, drafting SEO copy that I then edit into something a human wrote. It is fast at the boring eighty percent.
It is bad, still, at the parts that decide whether the project succeeds. It does not know your client's business. It will confidently build the wrong thing beautifully. It has no taste until you give it yours, sentence by sentence. It cannot tell you that a section should not exist. Left alone it produces plausible, generic, forgettable work, which is exactly the work that does not get anyone hired again. My job did not disappear. It moved up a level, to specifying, reviewing, cutting and owning the result.
The review step is not optional, and this is where experience compounds instead of expiring. I read what AI writes the way I would read a junior developer's pull request, because that is roughly what it is: fast, tireless, occasionally brilliant, and wrong in ways that look right. Spotting the wrongness that looks right is a skill you only get from years of shipping and maintaining real sites, which is why the fifteen years were not wasted. They were the prerequisite.
What this means for the people who hire me
The practical version is simple. I now ship production sites in four to six weeks instead of three months, and I can build a full web app solo instead of quoting it out of reach. Every build ships with AI-generated, human-edited SEO copy, a real performance budget, and analytics from day one. The price of a website did not need to stay high because the work was slow. It was slow because the tools were slow. The tools stopped being slow.
I am not romantic about any of this. WordPress still pays real bills and still wins for certain clients, which I will happily tell you to your face. But after fifteen years of doing it the careful, manual way, I made a deliberate choice to start over on purpose, because the alternative was to keep charging for time the work no longer takes. The judgement is still mine. The waiting is gone.
Want to see what AI-first looks like in practice? Browse the work, or tell me what you are trying to build.