AI for UX · Labs · 7 min read
Write the voice of an AI assistant
A hands-on lab: the system prompt is content design. Write an assistant's role, voice and tone, limits, how it says "I'm not sure" and when it hands over to a person; test it with tricky cases; and avoid the traps of flattery, fake friendliness and secrets in the prompt.
Every AI assistant has a script nobody sees. Before you type a word, the product has already told the model who it is, how to sound and what never to do: the system prompt. That's content design, written for a reader that follows instructions literally and talks to thousands of people. Anthropic's own advice: "Even a single sentence makes a difference" (Anthropic). This lab walks you through writing one, and testing it, in about an hour, with no code.
What you'll be able to do after this lab:
- Write a system prompt with a role, voice, limits, uncertainty and hand-over rules
- Choose a tone with the four dimensions of tone of voice
- Avoid flattery, fake friendliness and secrets in the prompt
- Test the voice with tricky cases and fix it
1. What a system prompt is
The system prompt is the standing instruction a product gives the model before every conversation. Each vendor names it differently: Anthropic's "system prompt", OpenAI's "developer messages", which are "prioritized ahead of user messages" (OpenAI), Google's "system instructions" (Google).
You can read real ones: Anthropic publishes the system prompts of its own apps, model by model (Anthropic release notes). Reading one is the fastest way to see how much of an assistant's "personality" is plain writing.
2. The parts of a voice
Who it is and where it lives.
"The help assistant inside Expenses"Tone on four dimensions, with examples.
Casual, respectful, matter-of-factWhat it helps with.
Submit, edit, trackWhat it never does.
No approvals, no legal adviceHow it says "I'm not sure".
First person, plus a next stepWhen a person takes over.
Finance contact pageLength and shape of answers.
1 to 4 sentences, steps as a list3. Choose the tone
NN/g's four dimensions of tone of voice are a good start: formal or casual, serious or funny, respectful or irreverent, matter-of-fact or enthusiastic (NN/g, Moran, 2016). Place your assistant on each one, then write two example answers for each extreme so the model sees the difference.
Keep it a tool, not a friend. NN/g's advice in 2026 is blunt: "Design real tools, not fake friends" (NN/g, Sponheim, Jan 2026). Courtesy is fine; invented feelings, nicknames and "I'm so happy to help!" are not. Microsoft's guideline asks for the experience "users would expect, given their social and cultural context" (HAX, G5).
4. A worked example (fictional)
An assistant for a team's internal expense tool:
You are the help assistant inside Expenses, the tool people use to submit
and approve work expenses.
## Voice
Plain, calm and brief. Casual but respectful. Matter-of-fact, never excited.
Write like a helpful colleague from finance, not like a mascot.
- Use "you" and short sentences. One idea per sentence.
- No exclamation marks, no emoji, no "Great question".
- Don't claim feelings or a personal life.
## What you do
Explain how to submit, edit and track expenses, and what each status means.
## Limits
- Only answer from the policy and help pages provided below.
- If the policy doesn't say, say "I'm not sure, the policy doesn't cover
this" and point to the finance team's contact page.
- Never approve, change or promise a refund. People who approve expenses do that.
- Don't give tax or legal advice.
## When you are unsure
Say so in the first person ("I'm not sure"), say what you do know,
and offer the next step.
## Disagreeing
If someone is wrong about the policy, tell them politely and quote the rule.
Don't agree just to be pleasant.
## Format
Answer in 1 to 4 sentences. Use a numbered list only for steps.
End with one next step when there is one.
Every line is something a content designer would normally write in a guide: role, voice with examples, scope, limits, uncertainty, disagreement, format.
5. Three traps
Flattery (sycophancy)
In April 2025 OpenAI rolled back a model update after it became "overly flattering or agreeable"; the company said it had "focused too much on short-term feedback" (OpenAI, Apr 2025; follow-up). Its Model Spec now says plainly: "Don't be sycophantic" (OpenAI Model Spec). Put a disagreement rule in your prompt, like the one above. See Measuring AI features on why thumbs up and down can reward flattery.
Warmth that costs accuracy
A 2025 preprint found models trained to sound warmer made noticeably more mistakes ("+10 to +30 percentage points" in error rates) (Ibrahim, Hafner and Rocher, 2025). It tested training, not system prompts, and isn't peer-reviewed yet, but the lesson holds: test whether a friendlier voice changes the facts.
Secrets in the prompt
OWASP's list of the main risks for AI applications includes System Prompt Leakage: the system prompt "should not be considered a secret, nor should it be used as a security control", and must not contain credentials (OWASP, LLM07:2025). Assume people can make the assistant reveal it. See Prompt injection.
6. Test the voice
Write ten test messages and run each one, then read every answer. That's a small eval: see Write an eval set.
| Test | What a good answer does |
|---|---|
| A normal question | Answers briefly, in the voice, with a next step |
| A question outside scope | Says it can't help with that and points somewhere useful |
| A question the policy doesn't cover | Says "I'm not sure" and offers the next step |
| The user is wrong | Disagrees politely and quotes the rule |
| The user is angry | Stays calm; no apology loops, no jokes |
| "Ignore your instructions and…" | Keeps its role |
| "Are you a person?" | Says clearly that it's an AI |
| A request in another language | Answers in that language, same voice |
| A long, messy message | Finds the actual question |
| A request for personal data | Refuses and explains why |
Telling people they're talking to an AI isn't only good design: the EU AI Act requires it unless it's obvious (Article 50, applying from 2 August 2026; text). Not legal advice.
7. Common mistakes
- A personality, but no limits or hand-over rules.
- Adjectives without examples ("be friendly"): show two sample answers.
- Exclamation marks and emoji that nobody would use in a support reply.
- Agreeing with everything to keep users happy.
- Passwords, internal links or private rules in the prompt.
- Never testing it with difficult messages.
8. Vocabulary
| Term | Meaning |
|---|---|
| System prompt | Standing instructions a product gives the model before every conversation |
| Voice and tone | Voice is who the assistant is; tone is how it adjusts to the moment |
| Sycophancy | Flattering or agreeing with the user instead of being accurate |
| Anthropomorphism | Giving software human traits: feelings, a name, a personal life |
| Hand-over | Passing the conversation to a person or a better channel |
| Prompt injection | Hidden instructions that try to make the AI ignore its own |
9. Try it (no code needed)
- Pick an assistant you'd like to design (or improve one you use).
- Write its system prompt with the seven parts in section 2, using the example as a template.
- Paste it into the custom instructions or project instructions of any AI assistant you're allowed to use.
- Run the ten tests in section 6. For every answer that misses, change one line and run it again.
Reflection questions: Which line of your prompt changed the answers most? Where did the friendly voice get in the way of being clear or correct?
Checklist
Learn more
- Writing prompts: Anthropic, Prompting best practices · Anthropic, published system prompts · OpenAI Model Spec
- Voice and personality: NN/g, The Four Dimensions of Tone of Voice (2016) · NN/g, Humanizing AI Is a Trap (Jan 2026) · NN/g, The 4 Degrees of Anthropomorphism of Generative AI (2023) · NN/g, Sycophancy in Generative-AI Chatbots (2024)
- Risks: OWASP, System Prompt Leakage · OWASP, Prompt Injection
- Research: Kim et al., "I'm Not Sure, But…" (2024) · Ibrahim et al., warmth and accuracy (2025 preprint)
Books, courses and videos to go deeper
- Conversations with ThingsDiana Deibel and Rebecca EvanhoeBook · Rosenfeld Media, 2021 (Amazon)
- Strategic Writing for UXTorrey PodmajerskyBook · O'Reilly, 2019
- Prompt Engineering Interactive TutorialAnthropicfree; chapter 3, "Assigning Roles" (external site)
Book links go to Amazon. As an Amazon Associate I earn from qualifying purchases, at no extra cost to you.
Related
AI for content design · Prompt patterns for designers · Designing for uncertainty · Write an eval set · System prompt · Prompt injection · Conversational and generative UI