Writing
Writing UX Microcopy with AI: Buttons, Errors, and Empty States
The smallest words in a product do the hardest work — a button label, an error message, an empty state. Here is how to write them with AI help so they sound human and do their job.

The smallest words in a product do the hardest work. A button label has about three words to tell someone what happens next. An error message has two sentences to rescue a moment that just went wrong. An empty state has one line to convince someone the screen isn’t broken.
And these are exactly the words AI gets wrong most often: it hands you “Submit” for a button, “An error occurred” for an error, and “No items found” for an empty state.
None of that is wrong, exactly. It’s just dead. And in a product, dead words aren’t a style problem — they’re a usability problem. People abandon forms over a confusing button. They close apps over a cryptic error. They assume a feature doesn’t exist because the empty state said nothing.
I’ve been writing microcopy with AI help for a while now, and the trick is the same as with any AI draft: the machine is fast at generating options and terrible at knowing what the words have to do. You supply the job. It supplies the variations. Here’s how I do it.
1. A button is a promise, not a label
“Submit” is a label. It describes the button. What a person actually wants to know is what happens when they press it — and “Submit” answers a question nobody asked.
Compare it with the promise version: “Send my application.” “Create my account.” “Download the report.” Each one tells the person the outcome, in their own words, before they commit to it. I’ve watched the same form convert better with “Send my request” than with “Submit,” because the second button removes a tiny moment of doubt.
The mistake is asking AI for “a button label for this form.” That prompt has no job in it, so the AI reaches for the most generic word it knows. Instead, I give it the promise: “Write 5 button labels for a form where a freelancer sends a project proposal to a client. Each label must say what happens after the click. Maximum 4 words. No ’Submit,’ no ’Click here.’”
Five options come back, I pick the one that sounds like something a person would actually say, and I test it on the real screen. “Send my proposal” usually beats “Send proposal” — the “my” makes it feel like their action, not the system’s.
2. An error message needs three parts — AI writes one
“An error occurred. Please try again.” That’s the most common error message on the internet, and it fails every test that matters. What happened? Nobody knows. Why? Nobody knows. What should I do next? Try again — the thing that just failed.
A useful error message has three parts: what happened, why it happened, and what to do next. “We couldn’t save your changes because your connection dropped. Check your connection and try saving again — your draft is still here.”
When I ask AI for error copy, I make the three parts explicit in the prompt: “Write an error message for a payment that failed because the card was declined. Say what happened, the most likely reason, and the one action to take next. Total under 30 words. No ’oops,’ no ’uh-oh,’ no ’something went wrong.’”
The “no oops” rule matters more than it sounds. AI loves chirpy error messages — “Oopsie! Something went sideways!” — because its training data is full of them. But a person who just lost twenty minutes of work doesn’t want chirpy. They want a way out. I keep a banned list for error copy: oops, uh-oh, whoops, went wrong, went sideways. If the AI uses one, I delete it and ask again.
3. Empty states are onboarding, not dead ends
The blank screen is the moment most products waste. A new user opens a project tool and sees “No projects yet.” Technically true. Completely useless. They came here to do something, and the product just told them there’s nothing to do.
The empty state is actually the best onboarding surface you have, because it’s the one moment you know exactly what the person wants: they want a first project, a first file, a first message. So write to that want: “No projects yet. Create your first one — it takes about two minutes.” Then put the button right there, in the empty state itself, not buried in a menu.
The prompt I use: “Write an empty-state message for a notes app with zero notes. One sentence naming what they’ll do first, one button label, one line saying how long it takes. Warm, plain, no exclamation marks.”
The exclamation-mark ban is deliberate. AI punctuates empty states like a children’s party — “Let’s get started!!!” — and it reads as nervous, not welcoming. A calm sentence beats an excited one when someone is figuring out a new tool.
4. Give AI the constraints, or it will write a paragraph
Microcopy lives in boxes. A button is maybe 180 pixels wide. A toast notification disappears in four seconds. An error banner has room for two lines on a phone, and the second line wraps badly.
AI doesn’t know any of this. Left alone, it writes button labels like “Download your free 14-day trial report now” — eleven words that will truncate or blow the layout apart — and error messages that run to a paragraph nobody reads while annoyed.
So the constraints go in the prompt, every time: character limits, word limits, and the screen the words sit on. “Button label, max 4 words, mobile width.” “Error banner, max 25 words, appears above a form on a phone.” “Tooltip, one sentence, max 12 words.”
I also tell it what sits next to the copy — a “Delete” button is fine on its own and terrifying next to “Save.” The words don’t exist in isolation, so I describe the neighbors or paste the whole flow in. Context is the difference between copy that fits and copy that fights the interface.
5. Test microcopy where it lives, not in a document
This is the step everyone skips, and it’s the one that catches the most problems. Copy that reads fine in a Google Doc can fail completely on the actual screen — because on the screen, it’s next to other copy, inside a layout, in the middle of a task, on a phone, while the user is impatient.
My test is simple: I put the words in the product (or a screenshot of it) and read the whole screen out loud, in order, as if I were the user. “Create account. Enter your email. Something went wrong. Try again.” Hearing it in sequence is what reveals the problems — the repeated words, the tone shifts, the button that promises one thing and the error that contradicts it.
I also test the worst case. “Your file is too large” means nothing when the file is 4.9 MB and the limit is 5 MB. Give the AI the numbers and it can write “Your file is 4.9 MB — the limit is 5 MB. Try a smaller file or a ZIP.”
6. Keep a microcopy voice list — and enforce it
A product’s voice is most visible in its smallest words. The button that says “Got it” and the error that says “An unexpected error has occurred” were clearly written by two different people. Users notice the inconsistency even if they can’t name it.
So I keep a short voice list for microcopy, and I paste it into every prompt: the words the product always uses (“save,” “send,” “your”), the words it never uses (“oops,” “submit,” “utilize,” “please be advised”), and the rule that error messages never blame the user. “You entered an invalid email” becomes “That email doesn’t look right — check for a typo.” Same information, no accusation.
The list is short on purpose — ten lines, not a style guide. And when the AI writes something off-brand, the fix isn’t to rewrite that one line. It’s to add one line to the list, so the next fifty lines come out right.
The honest truth
Microcopy is where AI’s weakness shows most clearly, because there’s nowhere to hide. A 1,000-word draft can absorb a few flat sentences. A three-word button can’t absorb anything — every word is load-bearing, and a generic word is a broken one.
That makes AI a brainstorming partner, not a writer. Give it the job, the constraints, the numbers, and the voice list, and it will hand you ten options in ten seconds — one of which is usually the right line with one word changed. That’s a good deal.
You have. So you decide the job, it drafts the options, and you keep the lines that survive contact with a real user. The smallest words in your product deserve the most deliberate process — because they’re the ones people actually read.
Frequently Asked Questions
Can AI write good microcopy on its own?
It can generate options, but it can’t know what the words have to do — the job, the constraints, the screen, the user’s state of mind. Give it the job, the character limit, and your voice list, and treat its output as a shortlist to choose from, not a finished line. The final call is always made on the real screen, by a human.
How short should a button label be?
Short enough to say what happens after the click, and no shorter. Two to four words is the usual range: “Send my proposal,” “Create account,” “Download the report.” If the label is one word — “Submit,” “OK,” “Continue” — it probably describes the button instead of promising the outcome. Add the outcome back in.
Should error messages apologize?
Briefly, if at all. “Sorry” costs words and doesn’t fix anything. One short apology is fine for a serious failure (“Sorry — we couldn’t save your work”), but never lead with it, and never apologize twice. The fix comes first.
How do I test microcopy without a designer?
Put the words on the real screen (or a screenshot of it) and read the whole flow out loud, in order, as if you were the user. Watch for repeated words, tone shifts, and promises the next screen doesn’t keep. Then test the worst case: the error with the numbers in it, the empty state a brand-new user sees, the button on a phone. If it survives that read, it ships.
Continue your workflow
- How to Write Landing Page Copy with AI Tools
- How to Write Better Headlines with AI (Without Clickbait)
Was this guide helpful?
Tell us what needs improving →
Your selection stays on this page. To send feedback to the publication, use our contact form.
Published 21 November 2026 · Text Craft Pro · AI-assisted educational writing. Editorial policy · Suggest a correction