Building This Website

An honest case study in early AI adoption — and the insight it returned to the Weavers

A Weavers report · the system examining its own construction · 2026

Why this report exists This website is not only a description of the Weavers — it is an element of the Weavers: part of the living system, subject to the same disciplines as everything else in it. One of those disciplines is that every encounter deposits what it teaches, including — especially — the failures. A record that keeps only successes is a highlight reel, not a curriculum. So this report tells the story of how the site was built, what worked, what went wrong, and the genuinely important insight the experience returned to the framework. In line with the Weavers’ philosophy of innovation and learning, the build itself was treated as an experiment in early AI adoption — and like most honest experiments, it taught more through its problems than its successes.

How the site was built

The site was built on WordPress.com, using the platform’s experimental AI assistant to create the initial site: the top-level index pages were produced by simply indicating what was required, and the assistant generated them. The lower-level pages were then created the traditional way — by copying text across from the Word documents in which the Weavers’ papers and guides already existed. The author brought some knowledge of HTML and CSS to the work: not expert, but enough, in ordinary circumstances, to fix a stubborn page by hand.

This mixed approach — AI-generated structure above, manually placed content below — was itself in keeping with the framework’s practice: use the new tools where they offer speed, keep the sovereign content in documents under the author’s own control, and treat the whole exercise as a learning encounter.

What worked

When everything worked, it worked well. The AI assistant produced the initial site quickly from plain-language instructions, and the results were excellent — pages that would have taken days of manual work appeared in minutes. The assistant’s open style allowed anything to be produced: there was little it would refuse to attempt, and much of what it attempted came out right first time. For getting from nothing to a working, presentable site at speed, the experience was genuinely impressive — and it is worth saying so plainly, because the honest account of the problems that follows is not a complaint about the tools. The speed and the capability are real.

What went wrong — and why it could not be fixed

The difficulty arrived, as difficulties do, through a minor problem. Page formatting on some of the manually created pages needed correcting — a small, ordinary job. The AI assistant fixed it. But the fix caused new issues on the AI-generated pages — and this is where the experience became instructive. Extensive AI analysis and multiple rounds of AI-attempted fixes failed to resolve the problems the AI’s own earlier work now exhibited. And the problems could not be solved manually either: the assistant had generated its pages through its own templates and structures, in its own way, and the result was a system that neither its creator (the AI) nor a human with working HTML and CSS knowledge could reliably repair.

Fortunately, the consequences here are trivial: some pages are less well formatted than they should be, and nothing more. The site works, the content is intact, and the reader loses only a little polish. But the pattern is anything but trivial, and it is the reason this report exists.

The insight returned to the Weavers

What this small experience demonstrates, at the safest possible scale, is a pattern the Weavers’ larger analyses have been warning about at national scale: AI can create systems that neither AI nor humans can subsequently fix — unless sound practical knowledge is maintained on the human side. The AI generated structures of its own design; when those structures misbehaved, the AI could not reliably diagnose its own artefacts, and the human could not repair by hand what had been machine-made in unfamiliar form. The repair capability that would once have existed — built through years of manual development practice — is exactly the capability that erodes when the manual practice stops. Use the AI for everything, and the knowledge that would let you fix what the AI cannot slowly disappears; and it disappears silently, because it is only ever tested on the day something breaks.

The same pattern, at every scale On this website, the consequence was imperfect formatting. But substitute the website for an enterprise platform, a public-service system, or a national data infrastructure, and the same pattern — AI-created structures that resist both AI and human repair, met by an organisation whose manual capability has quietly eroded — is precisely the dependency cascade the Weavers’ briefing papers describe as a serious national risk. The website was, in effect, a laboratory demonstration at a scale where failure costs nothing: the smallest possible version of the largest possible problem. The lesson is not “avoid the AI tools.” They built this site quickly and mostly well, and the Weavers itself is a testament to what AI-assisted work can produce. The lesson is the condition attached: keep the practical knowledge alive through practice — because the day the AI cannot fix its own work is the day the eroded manual capability is finally missed, and by then it cannot be quickly rebuilt.

What this deposits

In Weavers terms, the encounter has returned its insight, and the insight strengthens findings the framework already holds: the dependency cascade (capability delegated is capability that atrophies) here observed at first hand and in miniature; the gap between a system that works and a system that can be governed and repaired; and the growth loop’s hard corollary — that practical skills are maintained only by practising them, and that early adopters should deliberately preserve some manual practice precisely so the repair capability survives. The experience has been recorded honestly, per the framework’s own ledger discipline: what was tried, what it taught, and what it leaves behind for the next encounter.

There is also a quieter satisfaction in the story. The website that describes a framework about AI dependency and capability retention has itself, in its own construction, demonstrated the framework’s central warning — and then done what the framework prescribes: examined the failure, extracted the insight, deposited it, and carried on. The imperfectly formatted pages remain, for now, as small honest evidence that the system practises what it publishes.

A Weavers report on the construction of weaversai.co.uk. The site was built on WordPress.com using its experimental AI assistant for the top-level pages, with lower-level pages created from the Weavers’ existing Word documents. The account of what went wrong is deliberately honest, in keeping with the framework’s discipline that failures are curriculum. The insight it returns — that AI-created systems can resist both AI and human repair unless practical knowledge is kept alive through practice — has been deposited into the Weavers.