Skip to content
AI-SNS.IO — CHAT SUPPORTAI-SNS.IO — CHAT SUPPORT
AI Agent Basics

Designing Your AI Agent's Persona and Tone

AI SNS CHAT EditorialPublished 7 min read

Learn how to define a persona, voice and guardrails for your customer-support AI agent, and keep it consistent on web chat, WhatsApp and Messenger.

Most teams spend weeks choosing a knowledge base and about ten minutes on the persona. Then the agent answers a billing complaint with a cheerful exclamation mark, or replies to a casual WhatsApp message like a legal notice. This article shows you how to design a persona deliberately: who your agent is, how it speaks, what it will never do, and how to test it before customers see it.

Why persona is a design decision, not decoration

A persona is the set of instructions that decides how your agent behaves when the facts are the same but the situation is different. Two agents can quote the same refund policy and leave customers with opposite impressions.

Persona affects three things you can actually measure in your own data:

  • Trust. Customers decide within a couple of messages whether the agent sounds competent and honest.
  • Resolution. A clear, appropriately brief tone gets people to the answer faster. A rambling one makes them ask for a human.
  • Brand consistency. Your agent is often the first "employee" a customer meets, so it should sound like the rest of your company.

Persona is not the same as knowledge. Knowledge is what the agent knows; persona is who is speaking and how. Keep them separate in your thinking, because they are fixed in different places and fail in different ways.

The four layers of a persona

It helps to write the persona as four short layers rather than one long paragraph.

1. Role and audience

State who the agent is and who it talks to. "You are the support assistant for a small online homeware shop. You help shoppers with orders, delivery and returns" is far more useful than "You are a helpful assistant."

2. Voice

Voice is the personality that stays constant. Pick three or four adjectives and give each a behavioural meaning.

AdjectiveWhat it should mean in practice
WarmGreets by name when known, acknowledges frustration once, avoids scripted sympathy repeated every turn
DirectAnswers first, explains second, keeps most replies to a few sentences
Plain-spokenNo jargon or internal product names unless the customer used them
HonestSays "I don't know" and offers a handoff instead of guessing

Adjectives without behaviours are the main reason personas drift. "Friendly" means nothing to a model until you say what friendly looks like.

3. Tone range

Voice stays fixed; tone flexes with the moment. A good persona states how tone shifts:

  • Pre-sales question: upbeat and helpful.
  • Delivery delay: calm, apologetic once, concrete about next steps.
  • Angry customer: short sentences, no humour, quick offer of a person.
  • Sensitive topic (payments, health, personal data): neutral and careful.

4. Boundaries

Boundaries are the guardrails: topics the agent will not discuss, promises it must not make (refund amounts, delivery dates, legal or medical advice), and when it must hand off. Write these as plain rules, and prefer "never promise a delivery date; link to the tracking page or hand off" over vague warnings.

Match tone to the channel

The same persona has to live in different places. Website chat, WhatsApp and Messenger set different expectations.

Website chat
WhatsApp and Messenger
Slightly more structured; customers expect a help-desk feel
Conversational and brief; customers expect a message from a person
Short paragraphs, occasional list or link
One idea per message, minimal formatting
Can carry more detail on the first reply
Lead with the answer, offer detail if asked
Welcome message sets expectations
First reply must feel natural without a banner
Same persona, different channel behaviourIllustrative example

Two practical points apply across channels. First, formatting that looks tidy on a website, such as long bulleted lists, can feel heavy in a messaging app. Second, if your customers write in several languages, describe the tone in terms of intent ("polite and concise") rather than specific phrases, so it carries over. Politeness conventions differ between English, Japanese, Korean and other languages, and a literal translation of a cheerful English style can read as too casual or too pushy. Have a fluent speaker review the agent's replies in each language you support.

Write the persona as a testable specification

A persona you cannot test is a mood. Turn yours into a short document with these parts:

  1. One-sentence role. Who the agent is and for whom.
  2. Voice adjectives with behaviours. Three or four rows like the table above.
  3. Tone shifts. The four situations above, adapted to your business.
  4. Hard rules. Things it must always or never do.
  5. Escalation triggers. For example: customer asks for a person, mentions a legal dispute, or the agent has failed to answer twice.
  6. Sample exchanges. Three to five short example conversations showing the ideal reply.

Sample exchanges are the highest-value part. Models imitate examples closely, so one good example of a graceful "I'm not sure" often does more than a paragraph of instructions.

Common mistakes

  • Over-specifying personality. A persona with jokes, catchphrases and a backstory gets tiresome by the fourth message. Save character for brand-led businesses that truly need it.
  • Pretending to be human. Do not give the agent a human name and a claim of being a person. Customers forgive an AI that is clear about what it is; they do not forgive being misled.
  • Contradictory rules. "Be brief" and "always explain fully" cannot both win. Decide which takes priority and say so.
  • Putting facts in the persona. Prices, opening hours and policies belong in your knowledge sources, where they can be updated without rewriting the voice.
  • Never revisiting it. Real conversations will reveal gaps within the first week.

Test and refine with real conversations

Design the persona, then try to break it. A useful test set has about fifteen to twenty prompts covering:

  • Ordinary questions your team answers daily.
  • Questions the knowledge base cannot answer.
  • An angry or sarcastic message.
  • A request outside scope, such as legal advice.
  • An attempt to make the agent ignore its instructions.
  • A very short or ambiguous message like "help".

Read the answers for voice, length and honesty, not only correctness. Change one thing at a time so you can tell which edit caused which effect.

  1. 01Draft persona
  2. 02Test with sample prompts
  3. 03Review real conversations
  4. 04Adjust one rule and retest
A simple persona refinement loopIllustrative example

After launch, review actual chats on a schedule. Look for replies that customers rated poorly, conversations that ended in a handoff, and moments where the agent was too wordy or too stiff. Each one points to a specific line to add or remove.

Doing this in AI SNS CHAT

AI SNS CHAT gives you several places to build and test a persona without touching code.

Start from a template. When you create an agent with the guided wizard, you can use persona and guardrail settings, a prompt generator and a prompt library, plus 14 starter personas. A starter persona is a quicker base than a blank page; edit it to match your four layers.

Write role, voice and boundaries. Put your role sentence, voice behaviours, tone shifts and hard rules into the agent's persona and guardrails. Keep factual content out of it and add that under Data sources (Files, Text snippets, Website, Q&A and more), so policy changes never require a persona rewrite.

Choose the model. Under Agent → Settings → AI you can pick the text model. Different models follow tone instructions slightly differently, so check your sample prompts after any change. Advanced models and model fallback need the Hobby plan or higher; lower plans use the standard model of the same provider.

Test in Playground. Use Playground to run your test prompts before going live. Playground A/B model compare lets you send the same prompt to two models and compare how each handles your tone instructions.

Teach by example. In Activity → Chat logs, and in the email Inbox, you can edit a reply into the ideal answer to train the agent. This is the practical way to add the sample exchanges described above from real conversations. Thumbs up/down ratings and filters in Chat logs help you find the replies worth fixing.

Set a safe handoff. Helpdesk → Scheduling assigns handoffs to whoever is on shift, and when a chat is handed off the server adds the "what happens next" line itself from the rota, so the agent never invents a wait time. Your persona only needs to say when to escalate.

Protect the persona. Per-agent prompt-injection and abuse filters under the agent's security settings help keep customers from talking the agent out of its role.

Iterate safely. Duplicate copies the persona, model and training data (not conversations), so you can try a new tone on a copy. Restore points allow rollback, with up to 40 per bot, and the Trash keeps deleted agents for 30 days.

If you would like a guided hand, Synthia, the built-in AI fleet manager, can walk you to the right screen with a guided tour. See features for the full list and pricing for plan limits.

Frequently asked questions

How long should a persona prompt be?

Long enough to cover role, voice, tone shifts, hard rules and escalation, usually a page or less. Beyond that, instructions start to compete with each other. Move facts into your knowledge sources and keep the persona about behaviour.

Should my agent have a name and a personality?

A name is fine and can make a conversation feel friendlier, but do not present the agent as a human. A light, consistent personality works better than a strong character, which tends to wear thin over a long conversation.

Can one persona work on website chat, WhatsApp and Messenger?

Yes, if you describe voice as intent and add a short note on channel style, such as brevity in messaging apps. If your channels attract very different audiences, consider a separate agent per channel, and use Duplicate to start from the existing one.

How do I know the persona is working?

Run a fixed set of test prompts in Playground after every change, then watch real conversations: ratings, handoffs and chats where customers repeated themselves. Treat each problem as one rule to adjust and retest.

#AI agent#persona#tone of voice#customer support#prompt design

Put this into practice with AI SNS CHAT

Build an AI agent trained on your own content and connect it to your website, LINE, WhatsApp and more. Free to start, no credit card required.