Start with the feeling, not the website

Here’s the problem: most of us are terrible at describing feelings. “Warm.” “Premium.” “Clean.” Those words are so vague that AI just averages them straight back into the same generic output.

So this is a two-step process. The first prompt doesn’t build anything — it interviews you until the feeling is sharp enough to be useful. The second prompt turns that into an actual design direction you can hand to any build tool.

Write one messy sentence about how you want people to feel. Don’t polish it — bad first drafts are fine, that’s the whole point of step one.

Good messy starting points look like this: “I want it to feel like a really good independent bookshop.” · “Calm. Like nothing is being sold to me.” · “I want people to feel like they found something before everyone else did.” · “Serious but not corporate — my clients are surgeons and they hate fluff.” · “Like a letter from a friend who happens to be very good at this.”

That’s enough. Paste it into the bottom of the interview prompt below.

Copy this whole thing into Claude, ChatGPT, or whatever you use, and paste your messy sentence at the bottom. It won’t design anything yet — its only job is to interrogate the feeling until it’s specific enough to be actionable.

What to expect: it’ll ask you things you haven’t thought about. Answer quickly and instinctively rather than carefully — your gut is more reliable here than your reasoning. If a question feels impossible to answer, say so; that itself is information.

When you get the Feeling Brief back, read it once and check it for a flinch. If any line makes you go “no, not quite,” say so. One round of correction is usually all it takes. Then copy the whole brief for step two.

New chat, or the same one. Paste the Feeling Brief where marked. This prompt turns the feeling into a concrete visual system — actual fonts, actual hex values, actual easing curves — where every decision has to trace back to a line in your brief.

At the very end it hands you a single, self-contained build prompt (section 11) plus two alternative readings of your own brief, in case you picked the obvious interpretation of it.

Take the build prompt from section 11 and paste it into Claude Code, Lovable, v0, Cursor, or whatever you build in. Keep the full design direction open in another tab, because you’ll need it every time you add a page — new page, same system, no drift.

Feed it back in whenever the tool starts reverting to defaults. It will. Consistency is the thing AI build tools are worst at, and having the direction written down is what makes it fixable.

Vague prompts get averaged. When you say “make it feel premium,” the model has millions of examples of things labelled premium and returns the centre of that cloud — and that centre is, by definition, the most generic possible answer.

Specific constraints break the average. “Ink on uncoated paper, nothing gleams, everything sits still” isn’t the centre of any cloud. There’s only one place to go from there.

Skipping step one because you think you already know the feeling. You probably know it in your body and not in words — and the AI only has the words.

Answering the interview carefully. Answer fast. The considered answer is usually the safer, more generic one.

Accepting the first design direction because it sounds impressive. Read it and check whether it actually sounds like the thing in your head. If not, say what’s off — one correction, not five.

Using another website as your reference. You’ll get a version of that website. Use a room, an object, a book, a shop, a texture — something from outside the internet.

Letting the build tool drift. Three pages in, it quietly reverts to rounded corners and drop shadows. Paste the direction back in.