A Website You Instruct, Not Manage

Client context

Finlay Systems is a one-person AI consultancy. This case study documents its own production website, finlaysystems.com, not a client project. The site sells the work by being built the same way the work itself gets built. Finlay Systems also runs a second, older production site on Wix for its real estate media business, still on the old model, and a natural next candidate for the same rebuild.

The challenge

The old finlaysystems.com ran on WordPress, hosted on SiteGround. It didn’t look as professional or as polished as David wanted, and getting it there meant fighting the tool. Making almost any change meant finding the right widget, placing it by hand, and sometimes dropping into raw CSS to move something that wouldn’t sit where it should. Some changes just weren’t achievable without more patience than the change was worth, and got abandoned.

• Copy changes required a round trip: check the live site, open ChatGPT separately for the new wording, copy it back into WordPress, find the right field, replace it
• Anything visual, a new image block, a widget, a layout change, meant hunting through WordPress’s widget library and sometimes hand-editing CSS
• Bigger changes, a new page or a real layout change, could take hours, if not a full day
• No real preview step: a change either went live or didn’t happen

The cost of the problem

SiteGround’s large hosting plan cost $359.88 a year, excluding tax, renewing annually on September 16th. David is letting that renewal lapse instead of paying it again.

Time was the bigger cost. A plain text edit wasn’t the worst case, check the site, get new copy from ChatGPT, paste it into WordPress, find and replace the field. Slow, but tolerable. Anything structural, a new page, a new section, an image or widget placement, was a different story: hours of hunting through WordPress’s interface and occasional CSS, and building a real new page from scratch could eat a full day.

That cost was only manageable because David could absorb it himself. He’d built the old site on WordPress using his own prior web design experience, so the hours lost were his own time, not a bill. Most small business owners don’t have that background: for them, a widget that won’t cooperate or a layout change that needs CSS isn’t an afternoon of frustration, it’s a call to a freelance web designer or developer, typically billed by the hour or folded into a $500 to $1,500 a month retainer just to keep making ongoing changes.

The goal

Get structural changes, new pages, new sections, layout work, down from hours or days to minutes, on a site that actually looks as professional as the business behind it, without needing to hand-edit CSS or hunt through a widget library to get there.

I used to work at a company with a whole team dedicated to small website updates. Now one person can do it, with the right AI operating system behind them.

The existing workflow

Honest nuance: for a pure one-line text swap, the new system is only a little faster than the old ChatGPT-then-paste routine, that part was never the real bottleneck. The gap opens up on anything structural: a new page, a new section, a layout or design change, which is most of what actually kept the old site from growing.

The AI system

The site is built in Astro, version-controlled in Git, and hosted on Vercel, and it’s edited by talking to Claude Code the same way every other part of the business runs. David describes a change in plain language, sometimes pointing at a reference site whose visual style he likes and asking Claude Code to rebuild his own content in that style, or pulling a specific component from 21st.dev (the chat widget and several button styles on this site came from there). Claude Code plans and writes it, drafted directly in the codebase using the site’s existing design system, and the change runs on localhost first, a private, disposable copy of the real site, before anything ships.

A real example: every guide on this site, including the AI glossary for real estate teams, was built this way. David describes what he wants, less than 30 seconds of talking, then goes and does something else. The guide, structured for AI search visibility, is ready to review in a couple of minutes.

Implementation

• Astro as the site framework — compiles to fast, static pages with minimal JavaScript by default, a simple match for a marketing site that doesn’t need a heavier framework.
• Git / GitHub as the system of record — every change is a tracked, diffable, revertible commit, not an untracked live edit.
• Vercel for hosting and deploys, replacing SiteGround — free for a single project like this one; David’s own Vercel account runs $20 a month only because several unrelated projects now share it. Git-based deploys, preview URLs, and instant rollbacks replace an FTP upload and a locked page-builder interface.
• 21st.dev as a source for real, pre-built UI components and style references, pulled in and adapted rather than built from scratch every time.
• Claude Code as the orchestration and authoring layer, writing the code and drafting the copy in the site’s established brand voice, reusing existing design-system components rather than one-off styling. This is the system’s actual running cost: a $20-a-month Claude Code subscription.
• A required local preview step before any push — a running local copy of the real site, not a staging guess.

Human oversight and safeguards

• AI drafts, David approves. The AI writes copy and code from the instruction given, but nothing publishes without explicit human review.
• Localhost before live, every time. No change reaches finlaysystems.com without first being viewed and clicked through on a local preview.
• Git as a full audit and rollback trail. Every change is a commit; anything that ships can be traced and reverted.
• No auto-deploy without an explicit push. Publishing is a deliberate action, not a background process.

Results

David’s own working estimates from running both versions of the site, not a formal time-tracking study. The savings figure is a market-rate comparison, not a measured number, David’s actual cost is $20 a month for Claude Code, he built and runs the site himself, so nothing was literally “saved” out of pocket:

That $480+ figure is the conservative end of a $500 to $1,500 a month typical freelance or agency retainer for this pace of edits and new pages, minus David’s actual $20-a-month Claude Code subscription. Vercel hosting itself is free for a single site.

Business impact

The old site’s SiteGround bill and the hours lost hunting through WordPress’s widget library are both gone, replaced by a $20-a-month Claude Code subscription and a 30-second conversation, Vercel itself costs nothing for a single site like this one. Priced against the $500-to-$1,500-a-month a small business would typically pay a freelancer or agency retainer to publish and edit at this pace, that’s the real gap. That shift also changed what actually gets published: new guides and case studies now ship on an ongoing cadence instead of waiting for a rare redesign window, because writing one costs a conversation, not a day. For a consultancy that sells AI workflow automation, the site itself is now a working demonstration of the thing being sold, not a separate project still running on the old rules.

It's like having a web developer or a design team on call all the time. I just tell them what I want, a new page, a change, and they already know my business, how I communicate, and my site's style. What used to take days now takes minutes.

— David Finlay, Founder, Finlay Systems

What happened next

David has published more guides and case studies in the weeks since this migration than in the entire time the old WordPress site was live, feeding the AEO/SEO work documented in Finlay Systems’ other case study. The next candidate for the same rebuild is Finlay Systems’ second production site, currently on Wix for its real estate media business, not yet started.

What this means for other small businesses

The pattern here isn’t specific to a marketing site — it’s specific to any business whose website has become a bottleneck because editing it requires hunting through a page builder, fighting CSS, or paying someone else to make the change. A localhost-first, natural-language workflow removes the two biggest reasons small business websites stop growing: the cost of getting a change made, and the risk of a change going live untested. It’s honest, too, about where it doesn’t help much: a trivial text swap was never the real problem, structural change was, and that’s exactly where the gap is biggest.