On this page
The research articles on this website needed technical diagrams. At first that appeared to be a graphics problem: how could diagrams be generated consistently, kept in the right colours, exported at different sizes and inserted into WordPress without manual layout work?
We made progress down that route. A small generation process produced diagrams as image assets, including separate desktop and mobile variants. They were legible, and the production was repeatable. Yet something about the method still felt unnecessarily elaborate.
The question that changed the approach
Looking at the actual figures, I realised that almost everything in them was already a familiar part of a web page. They contained headings, short descriptions, bordered boxes, background colours and arrows connecting one stage to the next. None required detailed illustration or image manipulation.
Why produce a picture of a set of web elements when the browser could draw those elements directly?
That was the more useful engineering question. It changed the task from improving an image-generation process to removing the process entirely for this class of diagram.
Reusing what the website already knew
The site already had typography, colour tokens, spacing rules and light, dark and high-contrast presentation. A diagram rendered as HTML could inherit those decisions. Labels would remain selectable. Text could reflow naturally as the viewport narrowed. A single definition could work across appearance modes without maintaining separate image variants.
We replaced the graphical output with semantic figure markup and CSS. The content remained simple: a figure title, a short explanatory line, a series of labelled stages and a caption explaining the intended simplification. The page’s existing design system did the visual work.
The first results were better integrated into the article than the images had been. We subsequently refined the panel backgrounds, particularly in Light and Dark modes, without regenerating a single graphical asset.
The benefit was not merely appearance
The original approach had more moving parts: structured input, graphical rendering, multiple file outputs, insertion into WordPress and later maintenance of those assets. The web-native approach reduced the number of separate artefacts and made future revisions easier to reason about.
It also clarified an architectural boundary. The article owns the meaning of a diagram; the theme owns its appearance. That is the same separation used elsewhere in the site’s design: content describes what is being communicated, presentation decides how it should appear.
This does not mean CSS is the right tool for every visual explanation. A geographic map, a scaled scientific plot or a dense technical network might be better represented by a different medium. The point is to choose the simplest representation that can preserve the information accurately.
What I learned from the correction
The original solution wasn’t a failure. It demonstrated that the diagrams could be produced consistently, and it revealed the constraints that mattered. But it wasn’t the final answer.
This experience illustrates one of the benefits — and one of the obligations — of working with AI-assisted implementation. Plausible solutions can arrive very quickly. That makes it possible to test alternatives, but it also makes it easy to accept machinery that would not have survived a careful examination of the underlying problem.
The useful intervention was to question the premise, not simply ask for a better version of the same pipeline.
A smaller implementation emerged. It looked more at home on the site, required fewer separate assets and was easier to maintain. The improvement came from recognising what could be removed.
Connection to the wider research
This journal entry is one small incident within the development of johnstutely.com. The associated project study documents the broader website implementation, while the research programme examines the relationship between human decisions and AI-assisted engineering.
I don’t consider the lesson to be that AI always overcomplicates things, or that a human always sees the right answer immediately. In this case, we explored one route, inspected the result, challenged the assumption and replaced it with something simpler.
That willingness to reconsider a working implementation is worth preserving, regardless of how the code was produced.
Research context
Research programme: From intention to implementation: human-directed AI-assisted engineering
Topic: AI-assisted engineering
Explore further
Further reading is listed below.