How news from home became an evidence-led website in less than 48 hours

Why I built Save Jodrell Bank

I grew up near Jodrell Bank. With an interest in science, the Lovell Telescope always felt like more than a familiar Cheshire landmark: it was evidence that work with global significance could happen close to home.

In July 2026, I saw the news that future funding for e-MERLIN had not been prioritised. I kept seeing the complicated decision compressed into a much simpler claim: 'Jodrell Bank is closing'. The real story was more nuanced, but no less serious.

I wanted a place that explained the situation clearly, without disappearing into institutional jargon or exaggerating what had happened. I was on holiday at the time, so I worked in short bursts. Less than 48 hours after I started, Save Jodrell Bank was live.

That was elapsed time, not two sleepless days at a laptop. Still, it was much faster than I could have researched, planned, built and published the same project without these tools. This is the story of what I built, how generative AI helped, and where it still needed a human in charge.

The Lovell Telescope at Jodrell Bank Observatory - Benjamin Shaw, CC BY-SA 4.0. Resized for web display.

What had actually happened?

On 9 July 2026, UK Research and Innovation (UKRI) published its Science and Technology Facilities Council prioritisation outcomes for 2026 to 2030. The outcome for 'e-MERLIN + JIVE' says that e-MERLIN was 'not prioritised for further funding'. UKRI also says current Particle Physics, Astronomy and Nuclear Physics projects continue in full until their existing agreements end.

On 27 July, the University of Manchester responded. It said current e-MERLIN funding ends in March 2028, that Jodrell Bank remains open, and that it is working with partners to secure the future of the Lovell Telescope and e-MERLIN.

e-MERLIN is the UK's national radio astronomy facility. It links seven selectable telescope stations across up to 217 kilometres of Great Britain to Jodrell Bank by optical fibre. The Lovell or Mark II Telescope can provide the Jodrell Bank element, depending on the observation.

Jodrell Bank is broader than that one funding agreement. The Observatory and the separately managed Centre for Engagement - which leads visitor and heritage activity-have different roles. The independent Square Kilometre Array Observatory also has its global headquarters on the site. So no, closure of the entire site had not been announced. But without replacement support, the scientific future of e-MERLIN and the usual role of the Lovell Telescope beyond March 2028 were uncertain.

That difference wasn't pedantry. If I was going to ask people to care, the project had to be accurate about what had happened and honest about what nobody yet knew.

Why it mattered to me

Part of my connection to Jodrell Bank is simply place: I was born and raised nearby. But it is also about the science and heritage. Scientific work began there in 1945, the Mark I Telescope became operational in 1957, and the site joined the UNESCO World Heritage List in 2019. It is a living scientific landscape, not just a large dish preserved as a monument.

Then there is my work in digital research and technology at Airbus. Part of my role involves developing technology projects and working through UK research and innovation funding routes. Aerospace R&T and radio astronomy aren't the same funding environment, but in my experience both rely on specialist teams, equipment and accumulated knowledge. Once teams disperse and continuity is lost, that capability can be difficult to rebuild.

That gave the decision a professional dimension as well as a personal one. A line in a funding table can affect scientific capability long after the accounting period ends.

This remains a personal, independent project. My work provides context, not Airbus involvement or endorsement, and Save Jodrell Bank is not affiliated with UKRI, Jodrell Bank Observatory or the University of Manchester.

The idea

I decided to build an independent website that explains first and asks for action second.

I didn't want a conventional campaign page full of urgent slogans and an unexplained petition button. The site needed to tell readers what Jodrell Bank and e-MERLIN are, what UKRI had and hadn't decided, why the observatory matters, and which future outcomes remained uncertain. Only then should it offer proportionate ways to help.

The tone I aimed for was closer to a modern science publication or national museum than a political campaign. Every important date, figure and historical claim needed a route back to a source. Credibility wasn't a finishing touch. It was the product.

I registered savejodrellbank.com and savejodrellbank.co.uk through Cloudflare for a total of US$15.76, chose the .co.uk address as the canonical home, and redirected the other addresses to it. Those domains were the project's only infrastructure spend. Everything else had to be capable of running for £0 per month.

Research came first

It would have been easy to jump straight into building the homepage, but I deliberately started with the evidence. If the facts weren't right, the rest of the project didn't matter.

I used the deep-research tools in ChatGPT and Gemini to investigate the observatory's history, scientific work, funding, UNESCO status and the developing 2026 story. They helped me map the subject and surface material to investigate, but their compiled reports weren't evidence. I reopened the important primary pages and recorded their links, scope and access dates separately.

An AI-generated research report is a map of what to verify, not evidence in its own right. A fluent summary can merge two claims, drop a qualification or turn an estimate into an apparent fact. Using a second model doesn't automatically fix that; it can repeat the same error more confidently.

I kept the long reports as original working material, then used Codex to organise them into smaller Markdown files inside a /knowledge directory. That became the project's canonical internal research record. It records sources, uncertainty, conflicts and claims that still need checking rather than pretending every note is ready to publish.

The site's reviewed long-form articles live in a separate /content layer. That separation mattered immediately: while reviewing the reports, I found impressive-looking figures for funding, jobs and users that I couldn't establish from the primary sources available to me. I left them out. Preserving a number wasn't worth weakening the whole site's credibility.

Building it in small pieces

Rather than asking an AI to 'build a website', I first wrote down the decisions it shouldn't have to guess.

AGENTS.md defined the engineering and editorial rules. PROJECT_CONTEXT.md described the audiences, tone and product direction. Two phase plans turned the work into bounded milestones with deliverables, quality checks and stopping points. The durable context lived in the repository instead of one enormous conversation.

I then used a fresh Codex conversation for each milestone. The loop was deliberately repetitive: read the project rules and relevant knowledge, inspect the existing code, complete one defined task, run lint and a production build, review the changes, commit the finished unit with Git, then push it to GitHub.

The first phase built the foundation: visual system, typography, accessible navigation, responsive layouts, homepage, metadata and Cloudflare deployment. The second delivered 17 planned editorial pages, turning that landing page into a connected resource covering the observatory, its history and science, the funding situation, why it matters and ways to help.

Generative AI sped up the research, planning, coding and review, but only because the project had clear boundaries. Without those guardrails it would simply have produced mistakes faster. I still chose the objectives, checked the evidence, reviewed the interface and rejected weak output. A fast wrong answer would only have got the project into trouble sooner.

The plans also changed when the code gave me a reason. I originally expected the published articles to use MDX because the knowledge base was already Markdown. For this set of pages, typed TSX modules provided the structured callouts, images, headings and metadata I needed without another compiler layer. The knowledge stayed in Markdown; the reviewed publication layer became TSX.

Static by design

The finished site uses Next.js, React, TypeScript and Tailwind CSS, but the important decision is simpler: it is static.

Every editorial page is generated at build time and exported to out/. Cloudflare Workers Static Assets serves those files, and a push to the production branch triggers a new build and deployment. There is no database, CMS, long-running Node.js server or request-time page rendering.

That keeps the hosting target at £0 per month and leaves very little operational machinery to maintain. It is designed to be fast, easy to cache and light on client-side JavaScript. The architecture is intentionally simple. There are very few moving parts, which means there's very little that can fail unexpectedly.

What I learned

The biggest lesson was that technical and editorial verification are different jobs. Linting can catch a broken import and a production build can catch an invalid route. Neither can prove that a date, number or sentence is true. 'Does it build?' and 'Can I support this claim?' remained separate gates throughout.

The second was that AI can launder uncertainty. One model's 'reports suggest' can become another model's confident fact after a few rounds of summarising. Keeping research notes separate from public copy made those changes easier to spot. When I couldn't verify a claim, I qualified or removed it.

The third was that narrow tasks beat heroic prompts. One milestone per conversation produced smaller reviews, less context drift and clearer stopping points. It also made changing direction cheap when the implementation challenged an early assumption.

Most importantly, careful wording didn't weaken the case. Saying that Jodrell Bank remains open while explaining the genuine uncertainty around e-MERLIN was more credible than repeating 'Jodrell Bank is closing'. Accuracy and urgency aren't opposites.

What it became

Save Jodrell Bank is now a connected information resource rather than the campaign landing page I first imagined. It explains the observatory, its significance, the funding decision and the limits of what is currently known, then gives readers practical ways to help.

Generative AI made it possible to deliver that much work in less than 48 hours of elapsed time. It didn't remove the need for technical experience, primary sources or judgement. If anything, the speed made those controls more important.

For me, the result is more than a website. It let me use skills developed elsewhere to support one of the places that first sparked my interest in science.

Thank you for reading. You can visit the website here.