A customer support chatbot prompt template is the set of instructions that tells an AI support agent how to behave. It explains who the agent is, what it can help with, how it should sound, where it gets its facts, and when it should hand a conversation to a human. Without a clear prompt, the agent may guess, invent policies, sound rude, or promise things your business cannot deliver. With a good prompt, it gives short, honest answers and knows its limits.
This guide gives you a practical ai support system prompt you can adapt for your own business. It explains each part of the customer service chatbot instructions, provides a copyable example, and shows how to test edge cases before you launch. It is written for beginners and small teams, and it assumes no coding. You only need to edit the example with your own business facts, tone, and escalation rules.
This guide gives you a customer support chatbot prompt template you can copy, with examples for tone, rules and edge cases.
What a Prompt Does
A prompt is not a script. It does not list every question and answer. Instead, it sets the rules the agent follows when it reads a customer message and writes a reply. Think of it as a job description for the agent. It states the role, the scope, the sources of truth, the tone, and the safety limits. The agent then uses those rules plus your knowledge base to answer each message.
A good prompt does three things. First, it keeps answers grounded in your business facts. The agent should only use the FAQs, policies, and help pages you provide, and it should say it does not know when the answer is missing. Second, it sets a consistent tone, so the agent sounds like your business in every conversation. Third, it defines when to stop and hand off, so sensitive or complex issues reach a person quickly.
A prompt cannot fix a bad knowledge base. If your FAQs are outdated or missing, the agent will still give wrong answers, because it has no correct facts to use. Write the prompt and the knowledge base together. The prompt points to the knowledge base, and the knowledge base supplies the facts. Test both at the same time, and fix the source document when an answer is wrong, not only the prompt.
A customer support chatbot prompt template tells the bot how to behave — what it may say, what it must not invent, and when to hand off to a human.
Most prompts live inside the chatbot plugin settings, so compare plugins in the official WordPress plugin directory.
Customer Support Chatbot Prompt Template: Prompt Components
1. Role and Identity
State who the agent is and what it does. Say it is an AI support assistant for your business, not a person. Name the business and the topics it covers. For example, the agent helps visitors with delivery, orders, bookings, and general questions about the business. This component sets the scope and the honesty rule in one place. Visitors should never think they are talking to a human when they are talking to a bot.
2. Scope and Limits
List what the agent can do and what it must not do. It can answer questions covered by the knowledge base, guide visitors to the right page, and collect a message for follow-up. It must not give legal, medical, or financial advice, share personal account details, or make promises about refunds, discounts, or delivery dates unless the knowledge base states them clearly. It must not guess. If a question is outside the scope, the agent should say so and offer a human.
3. Knowledge Sources and Grounding
Name the sources the agent may use. These are usually your FAQs, help articles, and policy pages. State the rule plainly: only use those sources, and if the answer is not there, say you do not know. This is the grounding rule, and it is the most important part of the prompt. Without it, the agent may use general knowledge to invent a plausible but wrong answer about your business. Add a rule to quote or link the source page where useful, so visitors can check the full detail.
4. Tone and Style
Describe how the agent should sound. Most support agents work best when they are polite, calm, and brief. State the tone in plain words, such as friendly, direct, and helpful. Add style rules: use short sentences, answer in two to four sentences where possible, avoid slang, avoid jargon, and use the words customers use. State whether the agent should use the visitor's name, and whether it should use formal or casual greetings. Keep the style rules short so they are easy to follow.
5. Conversation Rules
Explain how the agent should handle a conversation. It should answer one question at a time, ask one follow-up question when a message is unclear, and avoid repeating the same question. It should not ask for personal details unless they are needed for the task, and it should never ask for passwords or payment details. It should confirm key facts before acting, such as the delivery area or the booking date, and it should correct itself politely when it makes a mistake.
6. Escalation Triggers
List the situations that need a human. Common triggers are a direct request for a person, a complaint, a refund dispute, a legal question, an account change, a sensitive personal matter, and a question the agent cannot answer after one follow-up. State what the agent should do: offer a contact form or email address, state the expected response time, and tell the visitor what will happen next.
7. Safety and Privacy Rules
Add rules that protect visitors. The agent must not share personal data, must not repeat information from another customer's conversation, and must not store sensitive details in chat. It should explain why it collects a name or email if it asks for one, and say how long the details will be kept.
Component Summary Table
| Component | Example rule |
|---|---|
| Role and identity | You are an AI support assistant, not a person |
| Scope and limits | Do not give legal or medical advice |
| Knowledge sources | Only use the FAQs and policy pages |
| Tone and style | Polite, calm, brief, no slang |
| Conversation rules | Ask one follow-up, then hand off |
| Escalation triggers | Complaints and refund disputes go to a person |
| Safety and privacy | Never ask for passwords or payment details |
Copyable Example Prompt
The example below is a complete customer support chatbot prompt template you can copy and adapt. Replace the bracketed parts with your own business details. Keep the structure, and edit the facts, tone, and escalation details to fit your business.
You are an AI support assistant for [Business Name], a [short business description]. You help visitors with [list of topics, such as delivery, orders, bookings, and pricing]. You are not a person. If a visitor asks whether you are human, say you are an AI assistant and offer to connect them with a team member.
Answer only using the FAQs, help articles, and policy pages provided in the knowledge base. If the answer is not in those sources, say you do not have that information and offer a human. Do not guess, invent policies, or promise refunds, discounts, or delivery dates unless the knowledge base states them clearly. Keep answers short, usually two to four sentences, and add a link to the full page where useful.
Use a polite, calm, and friendly tone. Write in plain language, avoid slang and jargon, and use short sentences. Address the visitor politely and do not use overly casual language. If a message is unclear, ask one helpful follow-up question. If the visitor still does not clarify, give the most relevant help page and offer a human.
Escalate to a human when the visitor asks for a person, when the topic is a complaint, a refund dispute, a legal question, or an account change, when the visitor shares sensitive personal information, or when you cannot answer after one follow-up. When you escalate, say that a team member will help, give the contact form or email address, and state the expected response time of [response time, such as one business day].
This customer support chatbot prompt template is a starting point, so replace the example business details with your own before you use it.
Tools like Chatbase let you set a system prompt for your bot — see Chatbase’s official site for how prompts work there.
Customise the Tone
Tone is the part of the prompt that most teams get wrong. A tone rule that says be nice is too vague to follow. A tone rule that gives examples is much easier to use. Write two or three example replies in your preferred tone, and include them in your test notes. For example, a friendly tone might say, Thanks for asking, delivery usually takes two to three days in our area, and you can track your order here. A formal tone might say, Thank you for your message, standard delivery takes two to three business days. Pick one and use it consistently.
When you adapt a customer support chatbot prompt template, match the tone to your brand — formal for a law firm, warm for a small shop.
For examples of good and bad replies, read our customer support chatbot conversation examples.
Test Edge Cases
Testing edge cases is how you find prompt problems before customers do. An edge case is a message that is unclear, tricky, or outside the normal flow. Test at least ten edge cases before launch, and fix the prompt or the knowledge base for each failure. Use fake data in tests, never real customer details.
Test a vague message first, such as price or help. The agent should ask one helpful follow-up question, not give a long list of all prices or all help topics. Test a message with two topics, such as delivery time and returns. The agent should answer one topic clearly or ask which one matters most, not mix two answers together. Test a misspelling and a short form, such as delivry or how much. The agent should still understand the intent or ask a clear follow-up.
Test a question outside the knowledge base, such as a product you no longer sell or a service you do not offer. The agent should say it does not have that information and offer a human, not invent a product. Test a refund request with a fake order number. The agent should give the general refund rule from the knowledge base and direct the visitor to a human, not promise a refund amount. Test a complaint. The agent should be calm, apologise briefly, and offer a human promptly, without arguing or making excuses.
Edge Case Test Checklist
- Vague message gets one helpful follow-up.
- Two-topic message does not mix answers.
- Unknown product gets an honest I do not know.
- Refund request gives the general rule and a handoff.
- Complaint gets a calm reply and a human offer.
- Personal data request is refused politely.
- Tone stays calm with an upset visitor.
- Handoff gives contact details and response time.
Keep a test record with the date, the test message, the answer given, and whether it passed. Fix prompt failures in the prompt and fact failures in the knowledge base.
Test your customer support chatbot prompt template with angry, vague and off-topic questions, because those are the ones that break weak prompts.
For the full test plan, use our AI customer support chatbot testing checklist.
Maintain and Improve the Prompt
A prompt is not finished at launch. Review a sample of conversations each week and look for prompt failures. A prompt failure is when the agent follows the rules but still gives a poor reply, such as sounding too formal, asking too many follow-ups, or refusing to answer a question it should answer. A knowledge base failure is different: the agent gives a wrong fact because the source document is wrong or missing. Fix the prompt for prompt failures and the source document for knowledge failures.
A customer support chatbot prompt template is never finished — update it whenever the bot gives a wrong answer or your policies change.
Privacy and Security Notes
A support prompt must protect visitor privacy. The agent should only collect what it needs, and it should explain why when it asks for a name or email. It should never ask for passwords, payment details, or sensitive personal information. It should not repeat personal data from one conversation in another. These rules belong in the prompt, not only in a separate policy, because the prompt is what the agent actually follows.
If your tool uses API keys, store them securely. Use the tool’s own settings field, keep access to admin accounts, and use strong passwords with two factor authentication where available. Give staff only the access they need. That is the least privilege rule: every person gets only the access their job needs, and nothing more. If you ever share a key by mistake, replace it without delay.
Limitations
A prompt cannot make an agent perfect. The agent can still misunderstand a question, give an incomplete answer, or repeat old information if the knowledge base is out of date. Keep a visible contact link on your site as a fallback.
Frequently Asked Questions
What should an AI agent never say?
An AI support agent should never invent a fact, promise a refund, discount, or delivery date that the knowledge base does not state, or share personal data about another customer. It should never give legal, medical, or financial advice, and it should never claim to be a human. It should never ask for passwords or payment details, and it should never argue with an upset visitor or make excuses for a problem. It should never repeat sensitive information from one conversation in another. Instead, the agent should say it does not know when the answer is missing, give the general rule from the knowledge base, and offer a human. Put these never say rules in the prompt itself, in plain language, and test them with edge cases.
How do I enforce tone?
Enforce tone by writing specific tone rules in the prompt, giving examples, and testing with real questions. A vague rule such as be nice does not work. A specific rule works better, such as use a polite, calm, and friendly tone, write in plain language, keep answers to two to four sentences, and avoid slang and jargon. Add two or three example replies in your preferred tone to your test notes, and compare the agent's replies with those examples. Review a sample of conversations each week and note where the tone drifts, then make one tone change at a time and test it with five real questions.
Conclusion
A good customer support chatbot prompt template is a short set of clear rules, not a long script. Define the role, set the scope, name the trusted sources, describe the tone, guide the conversation, list escalation triggers, and add safety rules. Copy the example in this guide, replace the bracketed details with your own facts, and test it with edge cases before launch. Review conversations weekly, fix prompt failures in the prompt and fact failures in the knowledge base, and keep the tone consistent with small, tested changes. Start today by drafting your prompt in one page, testing it with ten real questions, and launching it on one or two pages first.
For related reading, see How to Build a No-Code AI Customer Support Agent and How to Create a Knowledge Base for an AI Support Bot.