Gains:
- Ability to write button, error messages, idle status and onboarding texts with artificial intelligence in accordance with the brand voice and tone guide
- Ability to produce multiple variations of the same microtext and select them based on clarity, tone and length criteria
- Ability to audit AI text for inclusive language, accessibility and legal correctness
Every word in an interface is a design decision. UX writing is the task of writing micro texts such as button labels, error messages, empty status texts, onboarding screens and notifications, accurately and in line with the brand voice, without tiring the user. These texts are considered unimportant because they are short; whereas the difference between "Save" or "Save changes", "Something went wrong" or "Card number is missing" is the difference between the user completing the task and giving up. AI is extremely powerful at generating microtext variations; but you remain in control of tone, accuracy, and comprehensiveness.
Not without the brand voice and tone guide
By default, the AI writes in “average, neutral” language. However, every product has a voice — an enduring personality — and a tone that changes depending on the situation. A banking app can be reassuring and calm, a gaming app can be energetic and friendly. If you do not give this guide to artificial intelligence, it will produce texts that are foreign to your brand.
Practical method: give the model a short “voice and tone card” — 3-4 adjectives (“clear, warm, understated”), dos/don'ts (“don't use emojis”, “don't blame the user”) and 1-2 examples. This card is the common ground for all your microtext productions.
Tip: Write your voice and tone card once and paste it into each prompt. Consistency is best maintained with this card if many people write on the same product.
The formula for a good error message: what happened + why + next step
The most critical of microtexts is the error message, because the user is already angry. The AI's default output is often useless: "Something went wrong, try again." This message does not say what happened, the cause or the solution.
The good error message consists of three parts:
- What happened: In clear, non-technical language (“Payment could not be completed”).
- Reason (if known): ("Card number appears to be missing").
- Next step: Concrete action for the user (“Check the number and try again”).
And one ban: don't blame the user ("The number appears to be missing" instead of "You entered it wrong"). Improve the AI sketch with this formula.
Generate variations, select by criteria
The greatest power of artificial intelligence in UX writing is that it produces 5-10 variations of the same microtext within seconds. But this power is useless without criteria. Eliminate variations with three criteria:
- Clarity: Does the user understand it in one read?
- Tone: Does it fit the brand voice?
- Length: Does it fit on the button/area and does it not tire the eyes?
Micro text type
poor example
strong example
button
"send"
"Confirm appointment"
Error
"Incorrect transaction"
"The card number appears to be missing. Check and try again."
empty status
"The list is empty"
"You don't have any favorites yet. Save the products you like with the heart icon."
Approval
"Are you sure?"
"Should we cancel this order? This action is irreversible."
three mini cases
Case 1 — Button clarity increased conversion. One team extended the “Send” button to 8 variations with AI and selected “Confirm appointment.” In A/B testing (the experiment comparing two versions), the approval rate increased measurably; because the user saw what would happen on the button.
Case 2 — Incriminating message corrected. The artificial intelligence suggested "You entered an invalid email." The UX writer changed this to "The email address appears to be missing or incorrect." A small language change reduced the form abandonment rate. Lesson: blaming the user spoils the experience.
Case 3 — Legal error caught. “Free cancellation anytime,” the AI wrote for a subscription screen. However, the product's cancellation policy was different; this sentence was misleading and carried legal risk. The team matched the text with the actual policy. Lesson: microtext is also audited for legal correctness.
Vacancy and onboarding: making quiet moments speak
The first screen the user sees is often an empty state: no messages yet, no orders yet, no favorites yet. Inexperienced designers dismiss this screen as "The list is empty"; whereas the idle state is one of the most valuable moments that teach the user what to do. A good empty status text does three things: explains the status, states why it's empty, and invites the user to take the first action ("You don't have any favorites yet. Save products you like with the heart icon"). Onboarding texts undertake a similar task: they guide the user by teaching them one thing at a time, without overwhelming them. Artificial intelligence rapidly produces variations of these texts; your job is to keep them short, inviting and action-oriented, avoiding long explanations. Empty status is not a bug, it is an opportunity.
Tip: Each blank state is filled with the question "what should the user do when they come here for the first time?" Write with the question. The idle state is a hidden onboarding moment.
Copiable prompts
Sound and tone card: adjectives = <<clear, warm, plain>>; don't = <<emoji, blame, jargon>>; example sentence = <<...>>.Task: Suggest button text for "<<screen/action>>". Give 6 variations, write in one sentence next to each one why it has that tone. No more than <<X>> characters.
Write this error situation with the formula "what happened + why + next step": Status: <<error>>. Don't blame the user; Don't use technical jargon. Give 3 variations and mark the shortest, clearest one. Voice and tone card: <<...>>
Check these microtexts for inclusive language: are there gender assumptions, language that excludes people with disabilities, cultural assumptions, or unnecessary complicating language? List issues and fixes.Texts: <<list>>
Check this interface text for legal/factual correctness: Do claims like "free", "always", "guaranteed", "cancellable" match the product's actual policy ("<<policy>>")? Flag those that don't and suggest a safe alternative. Text: <<...>>
Weak prompt / Strong prompt
Weak: "Write an error message."
The result: a general, unresolved, toneless text like "Something went wrong."
Strong: "With this tone card, write 3 variations on the 'what happened + why + next step' formula for 'payment failed'; don't blame the user; tick the clearest one."
Result: Selectable variations that show solutions and match the brand voice.
Difference: strong prompt gives voice-tone card + formula + blame ban + variation.
Common mistakes
- Printing without giving a voice-tone guide. Brand personality disappears, texts become generic.
- Accusatory error language. "You entered it wrong" puts the user on the defensive.
- Failure to verify legal claims. Phrases like "free cancellation" may conflict with the actual policy.
- Skipping inclusivity. Texts that include assumptions of gender, culture, or competence exclude the user.
- Ignoring length. Beautiful text that does not fit on the button is useless.
In summary
UX writing is short but high impact design decisions. AI quickly produces multiple variations of the same microtext; The real value lies in screening them for clarity, tone, and length, refining error messages with the formula “what happened + why + next step,” avoiding blaming the user, and checking the text for comprehensiveness and legal correctness. Maintain consistency by adding the voice-tone card to each prompt; You make the final decision based on the product and the user.
Application task
- Write a tone card for your product that includes 4 adjectives and 3 "don'ts."
- Create 6 variations for a button with the first prompt and choose the best one based on the criteria.
- For an error situation, type "what happened + why + next step" message with the second prompt.
- Check all the texts you produce for comprehensiveness with the third prompt.
- Check a text containing a claim (free, guaranteed) for legal accuracy with the fourth prompt.
checklist
- [ ] I added the voice-tone card to each prompt.
- [ ] I eliminated variations by clarity, tone and length.
- [ ] I wrote the error messages with "what happened + why + next step".
- [ ] I did not leave any text accusing the user.
- [ ] I did an inclusive language check.
- [ ] I matched the assertive texts with the actual policy.