A few months ago I was halfway through a Saturday rebuild of a section of this site, and I caught myself doing something that would have sounded absurd a year earlier: I described the change I wanted out loud, in plain English, hit enter, and went to make coffee while it happened.
The change was waiting when I came back. Five files touched, all consistent, the dev server still green. I read it, adjusted one thing, and moved on. That small moment is the whole reason for this post.
The tool was Claude Code, and on its current model — Opus 4.8 — it’s become a fixed part of how I build. Here’s the honest account of why.
What it actually is
It’s worth being precise, because “AI coding tool” now covers a dozen very different things.
Claude Code isn’t autocomplete nudging me along a line at a time inside the editor. It’s an agent that lives in the terminal. It reads the entire project, not just the file I have open. I tell it what I want, and it plans the work, edits as many files as the task needs, runs the commands, reads the output, and fixes its own mistakes before handing the result back.
The closest everyday comparison is delegation. I’m not typing faster — I’m handing a well-described task to something that can carry it from start to finish while I think about the next one. Less a sharper pencil, more a capable collaborator who happens to have read every file in the repo and never tires of the dull parts.
The version that won me over
I’ve used every model in this line, and I don’t upgrade on press releases. I upgraded because the output got better in a way I could feel.
Opus 4.8 is Anthropic’s most capable model right now. The benchmark gains are real — it leads on SWE-bench Verified and a stack of other coding tests — but the number that actually changed my habits isn’t on any leaderboard. Anthropic reports it’s roughly four times less likely than the previous version to let a flaw in its own code pass unremarked.
In practice that means it second-guesses itself. It flags the part it’s unsure about. It tells me “this works but you should check X” instead of confidently handing me a quiet bug at the end of a long evening.
When you build alone, that’s everything. I’m the only reviewer. There’s no teammate to catch the thing I missed. A model that surfaces its own doubt is worth far more to me than one sitting two points higher on a chart and silent about what it got wrong. The faster, cheaper fast mode matters too, but mostly because it keeps the loop — ask, watch, correct, ship — tight enough that I stay in flow instead of drifting off mid-thought.
Why it suits the way I work
I build everything with Astro, for reasons I’ve written about at length, and the two turned out to fit together better than I expected — for a reason it took me a while to put into words.
An Astro project is just real files in real folders. Pages are folders. Posts are MDX files. Components are plain .astro and React files. Config sits right there in the open. There’s no hidden platform, no database to interrogate, no dashboard standing between me and the thing I want to change. That’s exactly the shape of project an agent thrives in — it can read the whole tree, see how a component is used before it touches it, and make a five-file change that stays consistent across all of them.
The same quality that makes Astro calm to work in — everything is a file you can see — is what makes it legible to an agent. And because I read every change before it lands, the perfect Lighthouse scores and Core Web Vitals I care about never slip. Speed stays a default, not something I have to win back afterwards.
When I built Astro Rocket, my own starter theme, having a partner that could hold the whole structure in view at once turned what would have been weeks of careful, repetitive scaffolding into a handful of focused days.
What it doesn’t do
Every post about these tools seems to skip this part, so I won’t.
Claude Code does not design anything. It has no taste, it doesn’t know who the site is for, and it will build the wrong thing beautifully if I brief it carelessly. The judgement is still entirely mine — what a page should say, how it should feel, which details earn their place and which are noise. Hand it a vague request and you get vague work back. Think clearly first, and it builds clearly. The thinking was always the actual job.
So the work stays mine. I decide the shape of things, I read every diff, and I’m the one accountable when something goes live. What’s gone is the friction between making a decision and seeing it real — the hour of boilerplate, the tedious refactor, the third pass to keep naming consistent across a dozen files. That was never the part I was proud of.
What changed for me
Strip it all back and the change is simple: I spend more of my time on the part that matters and less on the scaffolding around it. The layout, the typography, the copy, the small details that make a site feel like someone cared — that’s where the hours go now. Claude Code didn’t replace the craft. It cleared the clutter standing in front of it.
It hasn’t changed what I build either. Still fast, calm, well-made Astro sites — the same as ever. It’s changed how much of the work feels like building and how little feels like plumbing. When you put the whole thing together yourself, that shift is not a small one. It’s most of the reason the work still feels good.
If you’re curious, the best way to understand any of this is to try it on something tiny — open the Claude Code docs, point it at a small project, and watch how it works. That’s how it started for me.
And if you’d rather have a site built this way — carefully, by a person, with the right tools pointed at the right problems — that’s the kind of work I take on as a freelance Astro developer, based in Veghel.