Raian Pollock

Building our agentic website management

Five homepage designs side by side. The large left design is dark green like a chalkboard, with the headline "Your AI co-teacher for active learning.", a box to start a lesson, and a sample history lesson with printed handouts pinned below. Four smaller designs sit on the right in blue, white, editorial, and black and pink styles.
Five homepage designs for Flip Education, all built and compared in about a day. The large chalkboard design on the left is the one I chose.

For most of my career, how fast a company could change its website showed how fast the company could move. Changing a site meant weeks of waiting on shared specialists, and some changes we simply gave up on. At Flip Education, the AI product for lesson planning that I co-founded and ran with one other person and a team of AI agents, I removed the middle of that process. A complete homepage redesign went from request to review in three and a half hours.

Website · Flip Education, raianpollock.com, Solvue · Feb–Sep 2026

All work
  • 3½ hoursfrom my request to a complete, checked homepage design ready for review, against days or weeks before
  • about 3 weeksfrom the start of work to the live site for the first full overhaul, in April, covering the marketing pages, curriculum pages, emails and support chat
  • 1 personOne person changes the site through AI agents and the change goes live from one request. I estimate a normal company needs three or four people for the same change.

Closer to the person with the idea

AI models have collapsed this work onto the person who has the idea, giving sales and marketing teams far more control, most visible in experiments, user experience and product marketing. Newer models look at the work of older models and find many things to improve, so a website can be redesigned for the better as the models get better. One shared design system keeps the web pages, the emails, the printable lesson PDFs and the support chat window looking like one product. Between my idea and the live site there is no specialist, only the agents, the automatic checks and my review, then a count of what users do.

The challenge

Before, every change passed through people or platforms that were shared with other work. At ContentKing, a software company that watched websites for search problems, each site change went through the CEO and the developers who also built the product, so site work took hours away from the product itself. At SEOmonitor, which sells search analytics software to SEO agencies and where I was Chief Revenue and Operations Officer, changes went through marketing, but marketing could barely change the site because of the platform's limits. A normal change took days at least, most likely weeks. Once we tried to change the pricing page. The changes we wanted couldn't be made, so they were patched in with poor design. The new prices went live, but the page got worse.

Human decisions

  1. Could

    A CMS, the usual website editing tool, run by specialists, with outside designers and developers for any new template. Or an agentic CMS, a website editing tool with AI agents built in. But then you're still on their platform with its limits, it doesn't connect well with other technology you want to add to the site, and it gives you little freedom.

    Chose

    Code every part of the site from scratch with AI models, no CMS, so an idea goes straight from me to the live site.

    Because

    Every hand-off between the idea and the live site slowed us down. Remove the hand-offs and an idea goes straight to the site, with no platform's limits in the way, and the changes we used to give up on get made.

How a change reaches the site

I describe the change. AI agents then write the code for any part of the site. For a redesign they build several directions as complete pages, and I choose between real pages, not sketches. Before I see anything, scripts run automatic checks on every change, catching things like a factual claim with no source or English left on a page in another language, so my review goes to taste. I review for taste, direction and the polish of each detail. Every problem I catch in review becomes an automatic check that catches it the next time. After that the host builds and serves the whole site, and the blog and curriculum pages go out in the same build. Once a change is live, an analytics tool counts what users do there, visits, sign-ups and lessons started, and the next change starts from what users did.

A change moves outward from my idea through the agents, the checks and my review to the live site. What users do there starts the next one.

One afternoon

One afternoon in September I asked for experiments to raise the homepage and the other page types to a higher level. Three and a half hours later a complete homepage design was ready for my review that evening. Within about a day, five complete homepage designs had been built and compared. I went with the chalkboard and paper direction. Paper is what makes Flip Education different, because its lessons are printed and run mostly offline, and the other directions felt too techy. I loved how it came out.

Better models, better rounds

The design improved in rounds, because the models improved. We did a first full redesign in April across the marketing pages, the emails and the support chat window. In June we built three homepage designs in a day.

The same way on other sites

This site, raianpollock.com, works the same way. I asked for a full redesign one day around midday, four design directions were ready within the hour, and the new site went live the next day. Solvue is an AI product for IT help desks whose website I built. There, four design directions were built as full pages in a single afternoon, and the choice was made the same day.

A homepage design for Solvue, dark navy.A homepage design for Solvue, white with a green band across it.Solvue's live homepage, a light page with the headline "IT help that sees the screen."
Three stages of the Solvue homepage, the first direction built in August, the last of the four, and the homepage live on solvue.ai today. Side by side, they show how far the design work came.
The raianpollock.com homepage of July 2026, near-black at the top, "Most companies are adding AI tools. Almost none have redesigned how the company runs."The raianpollock.com homepage of September 2026, dark purple and the longest page, "... Few have redesigned how the company runs."A raianpollock.com homepage of September 2026, pale blue, "Operator. I build the systems a company runs on, then I run them."The raianpollock.com homepage, pale green, "I design and run GTM and Operations for the agentic era."
Four versions of the raianpollock.com homepage, oldest first. July, the first live page. September, the page live when I asked for the redesign. The direction I picked that day. The homepage live today. The last three span about a day.
The Barcelona address as a street map in the style of an old city plan, streets coloured by noise, with a score panel and a table of nearby places.The same address as a small 3D model of the streets, with the score of 92 in large blocky digits.The same address as a pencil and watercolour sketch, with the score in a hand-drawn circle.The same address as a grey hilly landscape, higher where it is more walkable, with the score in large thin numbers.
Walkably is my personal project. Type any address in 20 European countries and see how walkable it is, with the reasons behind the score. Here are four designs for the same Barcelona address. Map data by OpenStreetMap contributors. Noise data by Barcelona City Council.
The opening screen of Tá nos Números, with the title on grid paper, a timeline from 1995 to 2025, and the first question, Does one vote matter.The question More or fewer people in poverty and hunger, with 54.8 percent falling to 20.7 percent and a column for each year.The question How much Amazon forest was cleared, with 12,775 square kilometres in large type and bars for each year.
Tá nos Números is another personal project of mine. It puts every Brazilian government since 1995 side by side on the same numbers, such as poverty, hunger, jobs and Amazon forest. Every number comes from an official source and links to it.

What it would have cost

For a normal overhaul like this, the estimate from published pay and agency prices is three to six months of work redesigning and building the homepage and the main page types. An in-house team in Spain would run roughly €30,000 to €158,000 in employer cost, an agency about €35,000 to €132,000. The AI tools ran on a subscription, so the main cost was my own review time.

What I would keep, what I would change

I would keep the automatic checks and the one design system; they carry the quality. What I would change is my own expectation at the start. The call I got wrong was how much review the earlier models needed. In an earlier round, a few months back, a project I expected to be easy took more than a week of my own time. Now the models get close to right in one pass, so that cost has dropped a lot. It was still worth it.

What the setup needs

For most websites, two parts do the work. The first is an AI coding agent like Claude Code, the tool that writes and changes the site's code. The second is a host like Cloudflare, whose network serves pages to visitors around the world, using Cloudflare Pages, its service for hosting a website. This site runs exactly that way. Flip Education went with Vercel instead, because of how it serves pages and because the project is large. Getting there depends on where you start. An existing site on a website builder most likely has to migrate, meaning move off that platform into code, unless you're happy with the AI plugins it now offers. A brand-new site is super easy to start this way.

Advice for the team doing this

If a team runs a site this way, each team should change the part of the site it owns, as long as its changes can't break another part. The team that owns the website, usually marketing, gets access to all of it, and other teams get narrower access. Being found in search engines stays a shared concern, because a change made in one place can affect the rest of the site. Where I'd say don't do it: very large sites like a global online shop, sites where many people who don't work with AI agents edit pages daily, and anywhere every change needs a formal approval trail. For most startups and scale-ups, some form of this works, with rules on who may change which part, and the checks every team's change must pass.

Sources

Receipts. The Flip Education site is built with Next.js, a web framework, hosted on Vercel, and the code was written with Claude Code, Anthropic's AI coding agent. Key dates: the April 2026 release, the June 2026 round, and the September 2026 round.