Why we removed the dashes from every product we make

There is a mark that in recent years became synonymous with machine written text: the long dash in the middle of a sentence. Readers recognize it without knowing why, and they trust the text less. Our products, written partly with the help of models, were full of them.

We decided that no visible content, in any language, would carry a dash. Hyphens that belong to a name stay: an I-130 form is still an I-130.

A dash does different jobs and each one has a different replacement. An aside becomes parentheses. A short label followed by a dash asks for a colon. A long sentence splits in two. A numeric range is written with the word to. A title with a brand uses a vertical bar.

We wrote a script with those grammar rules and ran it over 325 interface strings, the schemas of 19 forms, email subjects, tab titles and 84 help center articles in EZMig. Every file was then reviewed by hand: the script proposes, a person decides.

With an AI assistant, changing the prompt is not enough: the model obeys most of the time, not always. We added a server side pass that strips dashes from every reply and every citation before they reach the client, with a test that guarantees it.

The result goes unnoticed, and that is the point. The text reads the way a careful person would write it, because a careful person reviewed it.