On this page
Questions you could ask me
Questions you could ask me
Could the same questions matter in your work?
Or describe your own challengeLook at a finished website and it is natural to imagine the work of a team. There is a visual identity, an editorial structure, carefully differentiated content, accessible interactions, working forms and the engineering needed to hold everything together. Those disciplines are often carried by different people, with decisions passing between them.
This website has brought those concerns together in a different way. I directed the work, made the design and architectural decisions, inspected the results and decided what needed to change. GPT contributed substantially to the implementation across those disciplines. I am grateful for that contribution, and I think it matters to describe it accurately.
What interests me is not the novelty of using an AI assistant. It is what becomes possible when the distance between an intention and a working result becomes much smaller, while responsibility for the outcome remains clear.
Work that usually crosses several desks
A website like this involves more than a design and some code. Someone must decide who it is for, how a visitor should navigate the material, how the evidence is organised, how the pages behave on different screens, what accessibility requires and how an enquiry should be handled responsibly. Those decisions affect one another.
In a conventional project, even a good team has to communicate those decisions across specialist boundaries. That collaboration can improve the result; different people see different risks. But a handover can also separate the reason for a decision from its implementation.
In this project I could hold the intention and inspect its implementation within the same working process. GPT helped explore and produce solutions across the frontend, application logic, diagrams, interaction patterns and test work. A decision in one area could be checked against the rest of the site without first becoming a separate brief for another discipline.
Direction is not the same as doing everything alone
There would be little value in pretending that I wrote every line or produced every candidate implementation unaided. I didn’t. GPT made a scale and range of practical work possible that would otherwise have required substantially more capacity.
Nor would it be accurate to describe the site as something the AI simply built for me. I brought the professional experience, objectives, constraints and judgement about what belonged. I challenged decisions that felt wrong, asked for alternatives, reviewed the running pages and required corrections when the behaviour did not match the intention.
That distinction is important. The assistant can contribute real implementation work without acquiring responsibility for the product. It can make a proposal convincing without making it correct. Someone still has to understand what the system is supposed to achieve and decide whether it has achieved it.
The running system kept us honest
Several of the most useful decisions came from looking at what was actually on screen, rather than accepting a plausible description of the result.
The editorial collection cards initially gave too much space to their illustrations. The artwork was distinctive, but it slowed down the work of finding something relevant. We retained the individual animated drawings and changed the shared cards so the titles and explanations led, with compact illustrations to the right. That was a design judgement informed by the real pages.
The navigation offered another example. Individual collections were accessible, but the main navigation highlighted different parent sections as readers moved between them. It was technically navigable and still conceptually confusing. The correction was not a new visual flourish: it was a clearer information architecture, with Work as the common home of the five collections.
The enquiry journey evolved in much the same way. A generic invitation became article-specific questions, then a contact form that could retain the chosen questions and prepare an editable starting message. Privacy information, administrative records, a visitor receipt and a simple first-party human check were then added to support the interaction responsibly.
Not every iteration improved the result. Some changes introduced layout regressions or required repair. The value of an assisted process depends on being willing to find those failures and correct them, rather than treating each generated change as progress by definition.
Experience becomes leverage
Much of my earlier work involved understanding a problem across the whole lifecycle: talking to people, finding the real constraints, choosing an architecture, building enough to test the decision and changing course when the evidence required it.
GPT has not made those years irrelevant. It has given that experience a different reach. I can investigate an implementation, compare it with the intention and move across disciplines without losing the thread of what the system is for.
That does not make specialists or teams obsolete. Collaboration, independent review, domain knowledge and operational responsibility remain essential in many settings. It does change the point at which a larger team becomes necessary, and what an experienced individual can establish before asking others to commit to a direction.
What the result shows — and what it does not
The result is visible here: a working website with connected editorial collections, original illustrations, an enquiry process and a growing body of research and project evidence. It is a practical example of one human directing substantial GPT-assisted work across several disciplines.
It is not a controlled productivity study, proof that a person can replace every specialist, or independent certification of the site’s accessibility, security or production readiness. Those claims would require different evidence. The implementation still has to be judged by its operation, testing and response from actual visitors.
There is enough in the result, though, to support a narrower and more interesting observation: the productive boundary between deciding what to build and producing a working version of it has changed.
A different development relationship
I used to think of a substantial implementation as something that necessarily moved through a sequence of disciplines, each translating part of the previous decision into its own form. Increasingly, I can direct a continuous process in which those decisions remain connected and visible.
For organisations, the opportunity is not simply to get more code. It is to examine difficult ideas, expose assumptions and arrive at something testable while the original problem is still close at hand. That can make the next conversation more informed, whether it leads to a prototype, an architectural review or a larger delivery team.
The lesson I take from this project is straightforward: GPT can contribute work across disciplines, and experienced human direction can keep that work coherent, accountable and useful. The technology is significant. So is the judgement required to put it to work.
Research context
Research programme: From intention to implementation: human-directed AI-assisted engineering
Topic: AI-assisted engineering
Explore further
Further reading is listed below.