Claude didn’t start my website redesign by asking what color I liked.
It started from the vault.
That matters because most AI-assisted design work begins in the wrong place. You give the model a vague brief, and it gives you something that looks like every other portfolio site: gradient background, soft cards, oversized hero claim, maybe an abstract blob in the corner. The result may be competent. It rarely feels inevitable.
For andreaguetti.com, the useful part was different. Claude had enough context from my notes to understand the taste behind the work: systems, writing, focus, maps, photography, knowledge graphs, notebooks, and the slightly obsessive need for things to connect. So the first direction didn’t feel random. Dark background, teal accent, dotted grid, mono labels, a large masthead. A structure closer to a field notebook or a Bullet Journal than a startup landing page.
The dotted grid was the part that made it click for me. I use systems because I don’t trust inspiration to show up on schedule. I like maps because they show relationships, and I like notebooks because they keep work grounded. The grid wasn’t decoration. It was a visual shorthand for how I think.
Claude didn’t get that from one prompt. It inferred it from the context I’d already been building.
Then Codex did what Codex is better at. It inspected the actual site, checked responsive behavior, audited accessibility risks, compared before and after, and made the boring fixes that make a design usable: navigation on small screens, heading hierarchy, tap targets, spacing, link behavior, deployment checks.
The same thing happened with draftandfocus.com. The WordPress theme had to inherit the same identity without becoming a copy. Draft & Focus is an editorial site. It needed to feel like a place for essays, not a personal calling card. So the shared language stayed (dark palette, the same teal accent, strong type, mono navigation, dotted background), but the implementation changed. Claude helped with direction and taste. Codex handled the child theme, CSS, templates, deployment, and the unglamorous QA pass across desktop and mobile.
That split worked. The Draft & Focus pass wasn’t a redesign for its own sake. The site already had a strong direction, so the work was mostly refinement: move post metadata where readers expect it, tighten the search page, improve mobile tap targets, fix the About page rhythm, check the Blog page in mobile landscape, add the missing social link in the footer. Small work. Real work.
This is the part of AI tooling that interests me more than the demos.
Not “write me a website.”
More like: read the context, infer a direction, make a first version, critique it against the actual experience, implement the fixes, verify what shipped.
The vault is the difference. Without it, the model only has the prompt. With it, the model has accumulated taste, recurring language, decisions, values, tools, projects, and contradictions. It can see that I care about systems but don’t want corporate productivity cosplay. It can see that I write about AI, focus, and photography, but don’t want a site that screams “AI founder.” It can see that I like structure, not clutter.
That doesn’t make the AI the designer, just a better collaborator. The judgment still had to stay mine. I said yes to the direction, rejected what felt off, asked for the mobile checks, pushed the theme into production, and decided when the work was done. The starting point was better because the context was mine.
AI gets more useful when it stops being a blank text box and works from a real body of notes. Not because the notes make it correct. They don’t. They give it something better than a brief.
A pattern to notice.
And sometimes that pattern is a dotted grid in the background.
