Twenty options is half the job: audition the finalists where they shipTwenty options is half the job: audition the finalists where they ship

A proper name can look perfectly respelled on the page and still come out of a voice model as three separate words. The pronunciation dictionary behind a daily audio show on this site respells names so the voice says them correctly, and its respellings split a name into chunks that the voice read as one word each, so a single name was spoken as three. The fix was an audition mode that renders several alternative respellings of the same name for me to choose between by ear, and one name needed three rounds before a respelling landed. Every candidate had looked right as text, which made the page the wrong place to judge them. For UX copy the same advice applies: the document you draft in is almost never the screen your customer reads, so the finalists belong in the real component before you pick one.

Generate widely, because you can't judge a single option

Asking a model for the button label gives you a button label, plausible and generic and a little too eager, while asking for twenty across three registers gives you a spread with two or three options you would not have thought of and a lot you would never ship. Choosing among them sharpens your sense of what the screen needs, which a single suggestion never does, because a lone option gives you nothing to compare it against. Tell the model what the screen is for, who is using it and what just happened to them, since a label for someone who has just finished a purchase and a label for someone whose payment has just failed are different writing problems that share a component. Then read the spread for what it repeats: if most of the options open on the same verb or land in the same cheerful register, you have found the model's default and can write the instruction that steers around it.

Carry three finalists, not one, into the audition

The audition mode worked because it put several candidates in front of me in the medium that mattered, and the equivalent for copy is a small test page that renders each finalist inside the real component, at the real width, in light and dark, next to the heading it has to follow and the field it has to fit. Three finalists per screen is enough to see differences without turning the review into a vote. Each one gets four passes: read it aloud in the order a person meets it, heading first and then the label; open it at the narrowest width you support and see whether it truncates or wraps into something that reads differently; listen to it through a screen reader, where a label that repeats the heading above it is spoken twice in a row; and put it in front of someone from the audience if it blocks a task or touches money.

The options that win on the page and lose in the audition are the ones with the most to teach you, so keep a note of which finalist lost and why. The note becomes the next instruction to the model.

Advice for the next screen

Keep the twenty options next to the one you shipped, with the reason each finalist lost beside it. A reason such as "wraps to two lines at 320 pixels" or "the screen reader says it twice" is the kind of finding a model can act on, so paste those reasons into the next request as the brief. As an illustration of the advice, a request for twenty labels for a failed-payment screen becomes sharper on the second round when it says: the heading already says the payment failed, so the button should name the action and leave the explanation to the heading.