How to Use ChatGPT for Website Redesign

Where a general assistant is genuinely strong

Thinking, before any design happens. Working out what your page should say, in what order, and to whom — that is a conversation, and it is one a capable assistant is good at.

It is also strong at critique. Paste your current copy and ask what a sceptical buyer would still be unsure about after reading it. Ask what the page is implicitly assuming the reader already knows. These produce useful answers because they draw on language understanding rather than visual judgement.

And it is good at generating the material a redesign needs: alternative headlines, FAQ questions people actually ask, section outlines, alt text.

Where it is workable but awkward

Producing markup. A general assistant will write you HTML and CSS, and for a single component it is often perfectly good.

Where it struggles is coherence across a whole page. Ask for a page in one go and you tend to get sections that were each designed independently — different spacing rhythms, inconsistent button treatments, a type scale that drifts. It is generating locally rather than applying a system.

The workaround is to constrain it: establish the type scale, spacing values and component styles first, state them explicitly, and then generate sections one at a time against those constraints. That produces markedly more consistent output, at the cost of your time.

Where it structurally cannot help

It cannot see your site. Unless you paste the content in or use a browsing tool, it is working from your description, which means the output is about a site like yours rather than your site.

That is the fundamental difference from a tool that takes a URL. Your product names come back approximated, your specific phrasing is paraphrased, and your page structure is invented rather than read. For a redesign specifically — where the point is to keep what you have and change how it looks — that is the wrong shape of input.

It also cannot see the result. It has no feedback loop on what it produced, so it cannot tell that the layout collapses at 900 pixels or that the contrast fails.

Prompting patterns that actually help

A few things reliably improve output, and none of them are magic phrases.

Give it constraints rather than adjectives. "Body text 16px, line height 1.5, spacing from a 4px scale, one accent colour" produces better results than "make it modern and clean", because the first is checkable and the second is not.

Give it your real content, not a description of your content. Paste the actual copy.

Ask for one thing at a time. Requests bundling layout, copy and styling changes together tend to get partially applied.

And ask it to explain its structural choices. Not because the explanation matters, but because it surfaces assumptions you can correct.

A sensible division of labour

The productive pattern for most people is to use a general assistant for the thinking and a purpose-built tool for the producing.

Work out what the page should say and in what order in conversation. Draft and sharpen the copy there. Then hand the actual page — as a URL — to something that reads it and generates the layout and markup, where consistency and responsive behaviour are handled by the system rather than by your prompting discipline.

Using one tool for both is possible and it means doing by hand the parts the other automates.