Microcopy that doesn't sound like a lawyer wrote it. Error states, empty states, CTAs, tooltips — generated with controllable tone, reading age, and accessibility-friendly phrasing. Three distinct variants per request, picked-recommended one highlighted.
Tell the model what kind of copy you need, where it sits, and the tone you want. It returns three distinct variants and picks the strongest default.
The smallest words on a screen do the heaviest lifting. The button label decides whether someone presses. The error message decides whether they recover or abandon. The empty state decides whether they understand what to do next. And yet microcopy is the thing teams most often hand to whoever's available, with no brief, at the end of a sprint.
Three things separate microcopy that ships well from microcopy that ships and gets quietly rewritten:
The tool exposes four tones because most microcopy decisions sit on this axis:
The Flesch-Kincaid grade level approximates the school grade a reader would need to understand a passage. Public-facing UK government writing aims for grade 9. The BBC News website averages grade 9–10. The Sun newspaper writes around grade 6.
For consumer products, grade 8–10 is usually right. Below grade 8 risks sounding patronising; above grade 12 actively excludes a meaningful share of your users — including time-pressed, distracted, multilingual, or low-vision users who all benefit from simpler structure regardless of their education level.
"You entered an invalid email" is technically correct. "We didn't recognise that email format" is the same information without the implicit accusation. The model is trained to avoid the first construction in error messages.
Generic CTAs ("Get started", "Continue", "Submit") are weak. They tell the user a button exists but nothing about what it does. The generator favours specific verbs ("Start with £5", "Open your first pot", "Begin in under a minute"). When you're given multiple variants, the picked default is usually the most specific one.
If you ask three writers for empty-state copy, you get three different angles — one focuses on what's missing, one on the next action, one on the user's likely emotion. The model is asked to do the same. If the variants come back too similar, regenerate.
Microcopy is rarely a one-shot job. Generate variants for error, empty, success and edge states, then pick a consistent voice across them.