Microcopy looks like the easiest thing to generate and is deceptively hard. It is short, so errors are invisible in isolation. It is constrained, so a character limit is easy to miss. And it is repeated across hundreds of screens, so a small inconsistency in how you say "delete" becomes a product-wide problem that nobody notices until it is everywhere.
Where AI clearly helps
- First-draft bulk generation. Twenty empty states, thirty error messages, a full set of button labels — produced in one sitting from a table of scenarios. The raw material is the win.
- Alternatives at a fixed length. "Five versions of this label under 20 characters" is a task models handle well and humans find tedious.
- Tone passes. Rewriting an existing set into a specified register — plainer, more direct, less apologetic — across many strings at once.
- Error message structure. What happened, why, what to do next. Models produce this three-part shape reliably once asked.
- Onboarding sequences. Step copy, with the constraint that each screen carries one idea.
- Tooltip and helper text. Short explanations, rewritten against a length limit.
- Accessibility-adjacent clarity. Plain-language rewrites of dense interface text.
Where it nearly got us in trouble
- Consistency across strings. Generate labels in batches and you get "Remove", "Delete", "Discard" and "Trash" for the same action. Models do not hold a terminology table unless you supply one and enforce it every time.
- Character limits. A generated string that fits the design in English may not in German, and may not after the button gets padding. Short copy breaks layouts silently.
- Error states that blame the user. Generated error text defaults to "You entered an invalid..." far more often than it should. Instruct it to describe the problem, not the person.
- Over-apology. "We're sorry, something went wrong" appears constantly and dilutes genuine apologies. Most error states need information, not regret.
- Consent and permission copy. What the user is agreeing to, and what happens if they decline. This has a legal dimension and should not be drafted by a model alone.
- Destructive-action wording. Confirmation text for delete, revoke or overwrite. A vague label here causes real data loss; be specific about what will be lost and whether it is recoverable.
Editing time, from our 30-day benchmark
Our 30-day test measured roughly 22 minutes of editing per 1,000 words on the low-cost dedicated writer against about 8 minutes on the general assistant — near $7 per piece of your time at $30 per hour. Microcopy is the exception where that gap narrows: at twenty characters per string, both tools produce usable raw material, and the real time goes into the terminology pass rather than into rewriting sentences.
A workflow that holds up
Write the terminology table first — the verb for each action, the noun for each object, the words you do not use — and paste it into every prompt, because this is the one thing a model will not maintain on its own. Generate in batches from a scenario list rather than one string at a time, so the batch reveals inconsistency instead of hiding it. Give explicit constraints: character count, tone, and who is being addressed. Then run a dedicated consistency pass over the whole output before anything reaches a design file — sort the strings alphabetically and read the verbs; that single step catches most of the drift. Keep consent, legal and destructive-action copy out of the generator.
Generate the volume, hand-enforce the vocabulary
Batch it, constrain it, then read the verbs. The terminology table is the deliverable, not the draft.
Try Claude →Frequently Asked Questions
Can AI write UX microcopy?
It produces usable raw material at volume — error states, empty states, labels, tooltips. The work it cannot do is keeping vocabulary consistent across a product, which needs a terminology table you enforce.
How do I keep generated strings consistent?
Maintain a terminology table with the verb for each action and the words you avoid, paste it into every prompt, and run a separate consistency pass over the finished batch before it reaches design.
Does it respect character limits?
Only if you state the limit and check the result. Short strings that fit in English often break the layout in translation or once button padding is applied.
Should consent or destructive-action copy be generated?
Draft it, then review it properly. Consent text has a legal dimension and destructive-action confirmations need to say exactly what will be lost and whether it is recoverable.