The constraints have changed

My company website has needed attention for years and I never got round to it. Two things finally forced the issue. Over the last six months the site's focus had moved from the case studies to the writing - what I am learning about technology and AI, and what I think businesses should do about it. But the design still said web agency. And I wanted to move from Jekyll to Hugo, because authoring and deployment are much quicker.

The speed and efficiency of the build was not the big surprise. What caught me out was how much of my ingrained caution turned out to be unnecessary.

The old trade-off

As a one-person business I have always built on frameworks like Bootstrap, Font Awesome and jQuery. Not out of affection - they provided a reliable way of making sure breakpoints, button sizes and accessibility were done properly. You traded big downloads for the peace of mind that a lot of things were handled by default. For twenty years that was a pragmatic compromise.

Not any more. The AI does the checking for you: it tests breakpoints, measures hit areas, runs Lighthouse and reports the numbers. The rebuild removed 92 Bootstrap files and every line of jQuery, and the site is quicker and more accessible without them.

The same shift affects scope. In the old days, after weeks of manual migration, I would not have had the time or the motivation to recolour and optimise the images from the old blog. This time it took a couple of hours. ChatGPT recoloured the images while keeping the detail. Claude wrote a shell script to turn the flabby PNGs into optimised AVIF and WebP. Coherent design, fast downloads, one afternoon.

Taste still matters

What the AI does not supply is a point of view. For months I kept a scrapbook of things I liked on other people's sites, such as the citations running down the right margin of this essay on deep learning. It still took me a while to settle on a theme. But once I landed on maritime charts, with images from around East Kent, most of the remaining decisions became easier. A chart has numbered sheets, landmarks drawn as engravings, and magenta reserved for cautions - all things that I used in the new design.

Give the AI no input of your own and your site looks like everyone else's. That is a useful thing to see once, because the stereotype is a map of what to avoid. But the choice of theme, the landmarks, the rule about what magenta is for - none of that came from a prompt. It came from months of noticing what I liked and having a firm opinion.

Start with a page both of you can read

I began with a design system written as a single HTML page, because both the AI and I could read it and make sense of it. With Claude, you can flesh out the whole design concept very quickly - colours, typography, layout. Then take a few key pages - in my case the homepage, a blog post and a case study - and iterate on static HTML prototypes of them. Get Claude to keep the design system page up to date as you go.

Once those pages are right, ask Claude to go through the old site and find every layout and module the new one will need: tables, charts, code blocks, footnotes, galleries. Have it mock them up and give you options to pick from - deciding between three options takes fifteen seconds, reworking the one it chose for you takes an afternoon.

Split the work into stages you can check

Then plan. I spent days iterating on the migration plan with Claude, had different models review it, and answered every question it raised before running anything. Keep a record of the edits - it is how the model learns to correct itself next time. The more time you spend up front, the less you spend correcting things after the code is migrated.

I split the migration itself into two phases. First a straight copy from Jekyll to Hugo with nothing changed. Claude could achieve that by screenshotting before and after and churning until they matched: 0% pages identical, then 57%, 92%, 95%, 100%. Then the redesign, on top of a new technical platform already proven to work. The redesign ran across eight stages, each in a fresh session, each with checks that had to pass before the next began. Splitting it into chunks like this meant I always felt in control of the process.

The reviewing was not as quick. Eight stages took under four hours of agent time, my checking took three days. An evaluation check proves the model did what the plan said. Whether the plan was right is a different question, and only a person can answer it. I wrote in April that the constraint moves downstream when the build gets fast, and I was the constraint.

And it is not the end of the world if you miss something. Afterwards I spotted plenty of things that needed changing. Before AI-assisted editing, you fixed that by hand across scores of pages, or maybe decided to live with it. Now you correct one page and ask Claude to apply the change everywhere else. Again, always making sure that the design system and other documents are kept up to date.

Be ambitious

What the AI will not tell you is when to stop. Claude is very keen on suggesting next steps ("build a suite of smoke tests"), and each idea it offers is another exciting rabbit-hole to disappear down. A site that can always be improved is a site that never quite ships. Deciding the design was done (for now) was a job I could not hand over.

What the AI gave me was permission. Twenty years of caution had taught me what a one-person business could not afford to attempt, and that list has just expired. I wanted some personal expression in the site, so ten photographs of local landmarks became engravings and icons at the click of a button. A seagull flies across the homepage. I would never have dared before. Hundreds of loads later, it still makes me chuckle.