IxDF Master Class

Better product UX copy with AI: prompt, review, and refine

Kalina Tyrkiel-SzymańskaSenior Content Designer at Zendesk

November 5, 2026

Hi, I'm Kalina

UX writer by trade, psychologist by training

Senior Content Designer at Zendesk, designing AI-powered experiences
10+ years in content design, UX writing, and strategic communication
Created Poland's first course on AI in UX writing
Trainer, university lecturer, and consultant, based in Kraków

Today's promise

You'll leave with a way to guide AI instead of fighting it, and to review copy so quality doesn't depend on who wrote it.

Poll

Which best describes you?

AI'm a content designer or UX writer

BI'm not a content designer, but we have content designers in our organization

CI don't work with a content designer

Poll

How do you use AI for UX copy today?

AI don't use it yet

BOccasionally, to get a few options

CRegularly, in my everyday design work

DMy team shares prompts, skills, or plugins

What you'll learn

Guide and evaluate AI copy

Use prompts and structured instructions to get clear interface copy, instead of writing every string yourself.

Keep copy consistent as a team

Set up a review process with designers and writers, so quality doesn't depend on who wrote the copy.

Use AI responsibly

Follow clear principles and your own judgment, and keep copy accurate and accessible.

Agenda

1The shift: from writing to guiding3 min

5Evals: how to test5 min

2What LLMs are good and bad at7 min

6Common issues3 min

3Generative vs evaluative tools4 min

7Workflows for your team6 min

4Structuring instructions7 min

8Principles and next steps3 min

Then Q&A. Includes 2 live demos.

1

The shift: from writing to guiding

From writer to editor

Before

You write every string from scratch
Quality depends on who wrote it
Guidelines live in a separate doc

After

You set the context and constraints
AI drafts, reviews, or offers options
Guidelines are built into the tools your team uses
You make the final call

Prompt, review, refine

1 · Prompt

Give context

Share the context, constraints, and a rough first draft.

→

2 · Review

Check the output

Compare it with your guidelines.

→

3 · Refine

Fix the instructions

Update the prompt or tool, so the next output is better too.

When the output is wrong, check what was missing from your instructions and add it. Everyone who uses them gets the fix.

2

What LLMs are good and bad at in UX writing

What LLMs do well

Following clear constraints

Voice guide, terminology, character limits, product facts

Keeping terms and format consistent

Select vs Choose, sentence case, date formats

Finding patterns across lots of copy

Inconsistent tone, structure, or formatting

Changing tone or complexity

Shorter, clearer, warmer, more formal, easier to translate

Critiquing and comparing options

Many options fast, plus feedback on your work

Following constraints and keeping copy consistent

Use	Avoid
Select	Choose, Pick
Sign in	Log in, Login
Save changes	Save Changes
Delete	Remove (when data is gone)

Example terminology list you can paste into instructions

LLMs follow a voice guide, glossary, or character limit when you provide it
Many default to Title Case. You can teach sentence case with a rule and examples
One tool can't catch everything. Focus on your key inconsistencies
Build specialized tools for UI elements your team writes again and again: error messages, empty states, notifications

Finding patterns and changing tone

Find patterns

Paste 50 strings from one flow and ask:

List inconsistent terms, capitalization, tone, and sentence structure. Group them by type and show where each one appears.

Useful before a release or a design system audit.

Change tone or complexity

Original

Your request could not be processed due to an unexpected error.

Shorter and clearer

We couldn't save your changes. Try again.

Easier to translate

Your changes were not saved. Try again.

Critiquing your work and comparing options

Input

Screenshot of the design with the copy, plus the stakeholder's comment

→

Ask

"Give me 3 options that address this feedback and explain the trade-offs."

→

You

Pick one, adapt it, and explain your choice in the reply

Only paste designs into a safe, approved enterprise AI tool.

It's you who makes the call.

What LLMs don't do well

Writing from scratch

Especially without context. The copy comes out generic and often doesn't fit your product.

Product context

Copy can sound plausible but describe something the product doesn't do.

Dependencies between products

It won't know another feature already has the same name.

Knowing what's clear to real users

Copy that seems simple to the model can still confuse people in the actual flow.

Always bring a rough first draft

No context

Prompt

Write an error message for a failed upload.

Output

Oops! Something went wrong. Please try again later.

Draft and context

Prompt

Draft: "upload failed, file too big". Limit is 25 MB. Inline, under the drop zone. Improve it.

Output

report.pdf is larger than 25 MB. Upload a smaller file or compress it first.

LLMs don't know your product or the products around it

Plausible but wrong

Don't worry, your changes are saved automatically.

This is wrong if the product doesn't autosave.

Hidden conflict

Introducing Workflows: automate your daily tasks.

Another product in your suite already has a feature called Workflows.

What helps: give product facts and system behavior, keep a glossary across products, and check facts with your PM and engineers.

Copy that seems clear to the model can still confuse your users

LLMs can critique copy and generate options, but they can't replace usability testing. Test copy with real users in the actual flow.

In the chat

Share one piece of AI-written copy that sounded right but was wrong

What did it get wrong? Product facts, tone, terminology, or something else?

3

Generative vs evaluative tools

Tools that generate and tools that evaluate

Generative

Creates copy

Drafts, variants, tone shifts, shorter versions, options to address feedback.

Evaluative

Checks copy

Compares copy against your rules: terms, case, length, tone, accessibility.

Tool type	Generate	Evaluate	Best for
Figma plugins and tools	Sometimes	Yes	In-place evaluation: copy in the context of the design and layout
LLMs, such as Claude skills	Yes	Yes	Generating and evaluating anywhere text lives: docs, strings, screenshots

Which tool to use when

Task	Type	Where
Reviewing a whole Figma file before handoff	Evaluative	Figma tool or plugin
Drafting options for an empty state	Generative	LLM with your voice guide
Checking an error message against guidelines	Evaluative	Claude skill or Figma tool
Auditing 200 strings exported from code	Evaluative	LLM with a specialized skill
Replying to stakeholder feedback	Generative	Approved enterprise LLM

Evaluate copy where it lives, for example in Figma. Generate copy where you can give the most context.

4

Structuring instructions

Why structure your instructions

For the model

It follows rules in sections and lists more reliably than rules in one long paragraph.

For your team

Anyone can read, review, and improve the same instructions.

For your tools

Markdown is the shared format for prompts, Claude skills, custom GPTs, and project rules.

Markdown basics

You type	You get	Use it in prompts for
`# Heading`	Main heading	Name of the tool or task
`## Section`	Subheading	Sections: Context, Rules, Examples, Output
`- item`	Bulleted list	Rules and constraints
`1. step`	Numbered list	Steps the model must follow in order
`bold`	Bold	Rules that must never be broken
`\| a \| b \|`	Table	Terminology lists and output format

What to include in a copy instruction

1

Task

What to do and for which UI element

2

Context

Users, product facts, where the copy appears

3

Rules

Voice, terminology, case, character limits

4

Examples

Good and bad, with a reason for each

5

Output

Format you want back: options, table, rationale

Example: an error message review skill

---
name: error-message-review
description: Review and rewrite error messages. Use when someone shares an error, toast, or inline validation.
---

# Error message reviewer

## Task
Review the error message. Then suggest up to 3 rewrites.

## Context
Ask for these if they're missing:
- Where it appears (toast, inline, modal)
- What caused the error and what the user can do next

## Rules
- Say what happened and what to do next.
- Use sentence case. Use "select", not "choose". Use "sign in", not "log in".
- **Never blame the user.** Avoid "invalid", "illegal", and "failed to".
- Toasts: 90 characters max.

## Examples
Bad: "Invalid input. Error 422."  Good: "Enter a date in the future, like 12/31/2026."

## Output
A table with: Issue | Why it matters | Suggested fix

Best practices for writing instructions

Be specific. Write "60 characters max", not "keep it short".
Explain why. The model applies a rule better when it knows the reason.
Show good and bad examples. Examples often work better than more rules.
Say what to do, not only what to avoid.

Give each tool one job. Use separate tools for errors, empty states, and notifications.
Ask for missing context. Tell the model to ask instead of guessing.
Keep it short. Remove rules it already follows.
Version and share it. Maintain it like any other design system asset.

Where to learn more

Markdown

Markdown Guide: basic syntax and cheat sheet
markdownguide.org

Prompting and skills

Anthropic docs: prompt engineering and Claude skills
docs.claude.com

OpenAI docs: prompting guide
platform.openai.com/docs

Start with what you already have

Most of your rules can come from your voice and tone guide, glossary, design system docs, and past review comments.

Live demo 1 · about 2 min

Generate and review copy with a Claude skill

Placeholder: demo tool and file

1

Open the error message review skill

→

2

Share a screenshot and a rough draft

→

3

Get a review table and 3 options

→

4

Pick, edit, and explain why

Watch for: does it ask for missing context? Does it follow the terminology rules?

5

Evals: how to test

What's an eval?

A repeatable test that checks whether your prompt or tool does what you expect across many examples.

1 · Collect

10–20 real test cases, including edge cases: long names, plurals, legal copy, tricky terms

→

2 · Define

Clear pass or fail criteria, written as a checklist

→

3 · Run and score

Score by hand, or use an LLM as a judge and spot-check it

→

4 · Compare

Change one thing, rerun, compare results

Example: an eval sheet

Test case	Sentence case	Approved terms	Fits limit	Next step	No blame	Result
File too large	✓	✓	✓	✓	✓	Pass
Session expired	✓	✗ "Log in"	✓	✓	✓	Fail
Wrong password	✗	✓	✓	✓	✗ "Invalid"	Fail
No internet	✓	✓	✗ 112 chars	✓	✓	Fail

For each failure, add a rule, an example, or more context. Then rerun the whole set.

Eval tips

Keep a reference set: strings you know are right, to compare against
Rerun after every change to instructions, and when the model updates
Log failures and turn each one into a rule or example

AI can check rules

Case, terms, length, words to avoid

People judge meaning

Tone, empathy, clarity, fit for the situation

Live demo 2 · about 2 min

Evaluate copy in place in Figma

Placeholder: Figma file and plugin or tool

1

Select a frame or flow

→

2

Run the evaluation tool

→

3

Review flagged issues in context

→

4

Fix and recheck

Watch for: what it catches in context that a text-only check would miss, like space, hierarchy, or repeated labels.

6

Common issues when generating copy with LLMs

Common issues and what helps

Issue	What helps
Title Case by default	A sentence case rule with examples
Generic filler: "Oops!", "seamlessly", "unlock the power of"	A list of words to avoid, plus examples in your voice
Too long: LLMs count characters poorly	Check length with a tool, ask for several lengths
Invented features or system behavior	Product facts, and ask it to flag assumptions
Terms change between runs	A glossary in the instructions
Meaning lost when shortening	Ask it to list what it removed
Over-friendly: exclamation marks and emoji in errors	Tone rules for each context
Accessibility gaps: "Click here", "see the panel on the right"	Accessibility rules, plus human review

7

AI-powered content design workflows

Workflow for content designers

Build

Turn guidelines into skills and evaluation tools

→

Draft

You and designers draft with the tools

→

Review

You review the copy that matters most

→

Improve

Run evals and update instructions

You build and maintain the tools, edit copy, and own its quality. Your guidelines get applied even when you're not reviewing.

Workflow for teams with a consulting content designer

Designers

Use shared tools for everyday copy: labels, tooltips, helper text
Run the evaluation tool before asking for review
Bring a draft and context to every review

Content designer

Builds and maintains the shared tools and rules
Reviews high-risk copy: errors, onboarding, pricing, legal, AI disclosures
Turns recurring review comments into new rules

Workflow for teams without content designers

Create one page with your voice, key terms, and formatting rules
Build one review tool for the UI element you write most
Peer review with a shared checklist

Borrow from your design system docs and public style guides
Test copy with real users in usability sessions
Bring in a content designer for an audit when you can

Without a content designer, evals and user testing matter even more.

Review checklist for any team

Clear: says what happened and what to do next
Accurate: matches what the product actually does
Consistent: approved terms, sentence case, formats
Concise: fits the space, no filler

Empathetic: right tone for the situation, no blame
Accessible: descriptive links, no direction or color-only cues, works with screen readers
Translatable: no idioms, no strings built from fragments

Use the same checklist for human review, AI evaluation tools, and evals.

8

Principles and next steps

Principles for using AI responsibly

You make the call

AI suggests options. You decide and you're responsible for the result.

Use approved tools

Don't paste confidential designs or customer data into public tools.

Bring context

Share a draft, product facts, and constraints.

Verify facts

Check system behavior with your PM and engineers.

Test with people

Only usability testing shows what users actually understand.

Check accessibility

Review every AI suggestion for accessibility before you use it.

What to try next week

Step 1

Pick one UI element

Choose the one your team writes most often: errors, empty states, or notifications.

→

Step 2

Write one instruction

One markdown page: task, context, rules, examples, output.

→

Step 3

Test and share it

Run it on 10 real strings, fix what fails, and share it with your team.

Thank you

Questions?

Kalina Tyrkiel-SzymańskaPlaceholder: LinkedIn or contact link