Rebuilding a website in 2 weeks with the help of AI
With a limited budget and no desire for a major redesign before 2028, we asked: What’s the smallest change that makes this site usable for its owners?
The Office of the Patient Safety Commissioner (PSC) improves the safety of medicines and medical devices by ensuring that patient voices are at the heart of the design and delivery of healthcare in England.
We support their website through our GovPress service. Although it’s a WordPress site, it was originally built with WPBakery – a proprietary page builder that embeds shortcodes directly into the content, locking in layout, spacing, and colour styles.
We’ve written about the pros and cons of site builder plug-ins before.
Enabling independent content management
This set up made simple edits, like publishing a new page, quite complicated and the team often needed some help. We saw an opportunity to build a simpler, more practical system that allows the team to manage their content independently.
With a limited budget and no desire for a major redesign before 2028, we asked: What’s the smallest change that makes this site usable for its owners? Our solution was to switch to Gutenberg, the standard WordPress block editor, using AI to make the transition affordable.
Validating the approach
We needed to quickly test if this approach would work. We conducted an AI-supported discovery phase, developing scripts to extract measurable data and provide evidence for PSC’s budgeting discussions.
Our scripts successfully identified 23 distinct shortcodes across 1,587 instances and 668 elements with custom CSS. We used AI to generate headings and evidence tables for the discovery report, while our team wrote the narrative. The final report clearly indicated the source of each section – script, AI, or human – ensuring transparency. Ultimately, it was AI-guided, not AI-written.
The AI-driven rebuild
We considered 2 conversion methods: reading shortcodes directly from the database or scraping external content. We chose the database route as it provided a more precise layout translation. (Though we also took the opportunity to look at a scraping approach for future projects where database access or source code might be unavailable.)
The conversion succeeded because we enabled the AI to assess its own work. By comparing screenshots of the live page against the new build, the AI generated an objective score and iteratively refined components. This loop involved a ‘doer’ agent to make edits and a ‘reviewer’ agent to catch regressions.
We shifted our focus from individual pages to a core set of components, styling each against its production equivalent to ensure consistency.
What we learned
AI does well on this kind of project when it has precise, automated feedback on its progress, but human oversight is essential to judge the final quality.
- AI needs concrete targets. Without clear criteria and self-validation, AI struggles to gauge its own success.
- Human oversight is vital. People can spot logical or visual errors in seconds that automated checks might miss.
- Code quality requires judgment. A page may look right, but only a human can ensure the underlying code is clean and maintainable.
- The editing experience is key. A component must not only function on the front-end, it also needs to provide an intuitive experience for content editors.
A streamlined manageable site
In 2 weeks, we moved from proving the concept to launching a live, reviewable site.
We removed 56,000 lines of legacy code and transitioned to a component-based system. The PSC team can now manage their own content easily, while visitors benefit from a more searchable website.