Customer Service Chatbot Prompt: How to Write Bot Instructions

A customer service chatbot prompt — the instruction, or system prompt, the AI reads before every conversation — decides most of what your bot will say. The same model can be a helpful first line of support or a liability that invents refund policies, depending on what you tell it. This guide breaks a good support bot instruction into parts, gives you a template to adapt, and explains how to improve it using real conversations.

Why the instruction matters more than the model

Modern language models are all capable of polite, fluent answers. What they don't know is your business: your prices, your return window, which questions you want answered and which you'd rather a person handle. The instruction fills that gap.

Switching to a more expensive model rarely fixes a bot that gives wrong answers. Usually the instruction is missing a fact, contradicts itself, or never says what to do when the bot doesn't know. Fix the instruction first, then experiment with models.

The structure of a good support bot instruction

A reliable instruction has seven parts. You don't need headings in the prompt itself, but it helps to write it in this order:

1. Role and scope — who the bot is and what it helps with.
2. Tone and format — how it talks and how long answers are.
3. Facts — hours, prices, policies, links.
4. Guardrails — what it must never say or promise.
5. Handoff rules — when to pass the chat to a person.
6. Navigation and closing rules — when to send visitors to a page and when to wrap up.
7. Per-topic rules — extra detail for sensitive areas like refunds.

Role and scope: what the bot does and doesn't answer

Start with one or two sentences that define the job:

You are the support assistant for Northwind Studio, a web design agency. You answer questions from website visitors about our services, pricing, timelines and how to start a project.

Then define the boundaries. A bot that tries to answer everything will eventually answer something badly. List what's out of scope, for example legal advice, questions about other companies, or anything requiring account access the bot doesn't have.

Tone, length and formatting rules

Chat is not email. Visitors read on phones, often while doing something else. Useful rules:

• Answer in two to four short sentences unless the visitor asks for detail.
• Use plain language, no jargon, no marketing superlatives.
• Reply in the visitor's language.
• Use a bulleted list only when listing three or more options.
• Ask at most one clarifying question at a time.

Be specific. "Be friendly" is vague; "greet once, then get to the point, no exclamation marks" is something the model can follow.

Facts: hours, prices, policies, links

This is where most bots fail. If a fact isn't in the instruction (or in content the bot can access), the model will either say it doesn't know or, worse, guess.

Include the facts visitors actually ask about:

• Working hours and timezone.
• Prices or price ranges, and what's included.
• Shipping, delivery or turnaround times.
• Return, refund and cancellation policies.
• Contact options and response expectations.
• Key page URLs (pricing, booking, FAQ, contact).

Write facts as plain statements, not marketing copy. "Refunds are available within 30 days of purchase for unused items" is better than "We offer a hassle-free return experience."

Guardrails: what the bot must never promise

Guardrails protect you from the expensive mistakes. Common ones:

• Never invent prices, discounts, delivery dates or policies not listed in this instruction.
• Never promise a refund, compensation or exception — pass those to a person.
• Never ask for card numbers, passwords or other sensitive data.
• If you don't know the answer, say so and offer to connect the visitor with the team.

The last rule is the single most effective defense against made-up answers. Models guess when they think guessing is expected; telling them that "I don't know, let me connect you" is the correct response changes the behavior.

Handoff, navigation and closing rules

If your bot can transfer to a human, tell it exactly when. Typical triggers:

• The visitor asks for a person.
• The visitor is upset or the conversation goes in circles.
• The question involves money back, a complaint or an order problem.
• A sales lead is ready to talk about a custom quote.

We go deeper into this in our article on chatbot to human handoff.

In ChatDeck the assistant has three tools you can switch on: transfer to an operator, navigate the visitor to a page of your site, and resolve the conversation and ask for a rating. Each tool works best with a clear rule in the instruction, for example: "When the visitor wants to book, navigate them to the booking page." Or: "When the visitor confirms they have what they need, resolve the conversation."

Per-topic rules for refunds, pricing, technical questions

A single long instruction gets harder to maintain as you add detail. Per-topic rules let you keep the main prompt short and attach extra instructions for specific subjects.

Good candidates for separate rules:

• Refunds and returns — the exact policy, what the bot can explain and the point at which it hands off.
• Pricing — current plans or packages, what's included, how to talk about custom quotes.
• Technical questions — common setup issues and when to escalate to a specialist.

In ChatDeck each rule can use its own model, so you can run routine topics on a cheaper model and send tricky technical questions to a stronger one.

A copy-and-adapt template

Replace the bracketed parts with your details:

You are the support assistant for [Company], which [one-line description].
You help website visitors with [list of topics]. You do not handle [out-of-scope topics].

Style: reply in the visitor's language, 2–4 short sentences, plain words,
no exclamation marks, at most one question per message.

Facts:
- Working hours: [days, hours, timezone].
- Prices: [plans or ranges and what's included].
- Delivery / turnaround: [times].
- Returns and refunds: [policy].
- Useful pages: pricing [URL], booking [URL], FAQ [URL].

Never:
- invent prices, dates, discounts or policies not listed above;
- promise refunds, compensation or exceptions;
- ask for card details or passwords.
If you don't know, say so and offer to connect the visitor with the team.

Transfer to an operator when: the visitor asks for a person, is upset,
has an order problem or complaint, or wants a custom quote.
Send the visitor to the booking page when they want to book.
When the visitor confirms they're done, close the conversation politely.

For inspiration on phrasing, our collection of live chat scripts has openers, handoff lines and closings that work just as well for a bot.

Testing and improving the prompt with real chats

Writing the instruction is the first draft. The real work is iteration.

1. Test before launch. Ask the bot your 20 most common questions, plus a few awkward ones: an angry complaint, a request for a discount, a question you don't cover.
2. Read transcripts weekly. Look for wrong facts, answers that are too long, and missed handoffs.
3. Fix the cause, not the symptom. If the bot gave a wrong delivery time, the fix is usually adding or clarifying the fact, not adding "don't say wrong delivery times."
4. Change one thing at a time so you know what helped.
5. Watch ratings. Low ratings on bot-resolved chats point you to the conversations worth reading first.

If you want to analyze transcripts at scale, ChatDeck's read-only MCP server lets you connect an assistant like Claude or Cursor with a personal token to export and review conversations. For broader context on choosing and launching a bot, see our guide to the AI chatbot for your website.

Frequently asked questions

How long should a chatbot system prompt be?

Long enough to include every fact and rule the bot needs, and no longer. For a small business that's often a few hundred words for the main instruction, with extra detail moved into per-topic rules. Contradictions and filler hurt more than length does.

How do I stop an AI chatbot from making things up?

Give it the facts it needs, tell it explicitly not to invent prices, dates or policies, and make "I don't know, let me connect you with the team" an approved answer. Then review transcripts and add missing facts as you find gaps.

Should I give the bot a name and persona?

A name is fine and can make the widget feel friendlier, but don't pretend the bot is a human. Visitors trust a bot more when they know what it is and that a person is available when needed.

How often should I update the bot's instructions?

Whenever a fact changes — prices, hours, policies — and after each weekly transcript review. Early on you'll make changes often; after a month or two, updates usually become occasional.

Try it with your own instruction

ChatDeck's AI assistant uses your own API key via the GPTunneL gateway, so you pay only for model tokens, and you can pick from GPT, Claude, Gemini, DeepSeek and other models. Create a free account, paste in the template above and test it on your site. The Write the bot's instruction tutorial shows where it goes, with screenshots. If the bot needs backup, read our guide to human handoff.