Three rules produced a site the client was pleased with in an evening: ground every option in the client’s world, give every option an honest tradeoff, and never rename an option between rounds.
Ask an AI tool to design a website and you get the same site everyone else got: a cream background, a serif headline, a terracotta button. Or the dark one with the acid-green accent. These are not choices. They are the statistical centre of what the model has seen. And ask a non-designer to choose between three designs and you get “the middle one”, which is fine until you need to know which one that was.
Ground each option in the client’s world
The brief was a small firm offering technology leadership to businesses too small to employ it. The three first-round options each took one thing from that world and made it the signature: the fixed-price proposal sheet an owner actually signs, stamp and all; the client’s own org chart with the CIO seat drawn empty; the whole page as an operations status board. One dark, two light. None of them could have been produced for a different brief, which is the test.
Give every option an honest tradeoff
Each option went on the canvas with a sticky note: what motivated it, and what is wrong with it. The dark one “reads senior and expensive”. The status board is “techy; copy must stay in the owner’s language”. A set where only the designer’s favourite gets a case made for it is a rigged vote, and clients can tell.
Run a critic that has not seen the work
Before the client saw anything, a fresh-context reviewer, given only the artboard sources and told to name the tells of AI-generated design, went through every option. It found real problems each round: frames shorter than their content, a stamp covering table text, numbers I had invented for realism, and a warning that dark navy plus brass plus serif is the stock private-equity look. All fixed before the client saw it. A designer reviewing their own work sees the rationale; a stranger sees the page.
Keep the names stable
The client asked whether light or dark suits this kind of site. I said light: a business-to-business services site for owners is a category with conventions, long copy reads better on light, and the near-black-with-accent look is the current default that will date fastest.
Round two was three light directions on a second page of the canvas, named D, E and F, with A to C left where they were. The client said “the middle one in the second batch”, and because nothing had been renamed or reshuffled, that was answerable. It was E, the roadmap. I had expected them to pick the org chart.
Then the one I did not expect, again
The roadmap went live. A few hours later the client said it looked a bit too corporate, and asked for three more interesting options with one constraint: same copy, same fonts.
That constraint turned out to be the best brief of the day. With the words and type fixed, the options could only differ in composition and material, not decoration: an engineering blueprint, a leadership-meeting whiteboard, a bold colour-block calendar. They picked the whiteboard, the warmest and least corporate of the three. It was built into the site the same evening, because the tokens had been designed to carry a change like that.
The lessons I keep
- Every option gets a tradeoff written down.
- Option letters never change across rounds.
- A critic with no memory of the process finds what the process hid.
- When a client names a constraint, take it literally. It is usually doing more work than it looks.