Lovable
Agent, preview, and host in one place. I describe what I want in plain English, the page updates in front of me, and I keep going until it looks right. Publishing is a button instead of an afternoon.
meta
How I built this
This site was built with Lovable, Claude Code, and Codex. I wrote the brief and the copy, the agents wrote most of the code, and I said no to a lot of what they offered. Here is how it actually went.
LovableClaude CodeCodex
The tools
Claude Code
For the work that has to reach into the files themselves: pulling every piece of copy out of the components and into one content file, wiring the Calendly lookup so the API key never touches the browser, and tracking down the single style rule behind a bright line on the edge of a tile.
Codex
Used to take another run at the same problem. Different model, different instincts. When both agents land on the same approach I move fast. When they don’t, that disagreement is usually the real design question.
How it came together
The brief comes before the code
I wrote down who the site is for, what a recruiter should understand in the first ten seconds, and what I refuse to sound like. That document is why the build moved quickly. A vague prompt gets you a generic site in half the time.
One content file, every component reading from it
All copy, links, and image references live in a single file. Components only read from it. Changing a headline, a case study figure, or a book blurb never means touching the design, and an agent can’t quietly rewrite a number while it’s refactoring something else.
Structured content, not decoration
Case studies are delivery steps, result groups, and lessons. Photos are captions with places attached. Each news edition is a list of stories with sources. That structure is why a page can look completely different without being rebuilt.
Automations that cost nothing to keep
The savings calculator, the role fit scorecard, the time zone conversion on coffee times, and the QR code all run in the visitor’s browser. No model calls, no metered service, no monthly bill. The only outside data is my Calendly availability, fetched on the server and cached for a minute.
Where AI deliberately isn’t used
The skills match is rule-based keyword matching, not AI. The savings calculator is plain math. You can see exactly why you got a result. No black box when the stakes are your time.
The slow part is taste
None of this was hard to build. What took time was deciding: which photo caption was wrong, which impact figure was a team result and had to say so, and the news masthead that arrived in ornate lettering and went right back out again. That judgment is the job, and no agent hands it to you.
Why it’s set up this way
Good product judgment means picking the simplest tool that works. Sometimes that’s a model. Often it’s a lookup table. This site is small, fast, cheap to run, and easy to change, which is the same test I bring to any product.