System Prompt Examples for a Business Chatbot

A system prompt is the written set of rules your AI assistant follows in every conversation. A good one for a business chatbot is short and specific. It says who the assistant is, which sources it may use, what it must never do, how it sounds, when it hands over to a person and how its answers look. Clear limits beat long prompts.

Below are the parts, two copyable examples (a shop and a service business) and the ways rules usually fail.

Key takeaways

• A system prompt is your rulebook. Customers never see it, but the assistant reads it every time.
• Six parts cover most small businesses: role, sources, limits, tone, handover, format.
• Keep facts such as prices and hours out of the prompt. Put them in a source the assistant reads.
• Give the assistant an exact sentence for the moments it cannot help.
• A prompt is not a guarantee. Test every rule with questions built to break it.

What is a system prompt?

A system prompt (also called developer instructions or custom instructions) is text sent to the model before the customer's first message. It sets the behaviour for the whole conversation. The customer's own messages are the user prompt.

The major providers describe the same split. OpenAI's prompt engineering guide says developer messages carry the application's rules and user messages carry the end user's input, with developer messages taking priority. Google's Gemini prompting strategies advise placing "essential behavioral constraints, role definitions (persona), and output format requirements" in the system instruction. Anthropic's prompting best practices say that setting a role in the system prompt focuses the model's behaviour and tone, and that even one sentence makes a difference.

Custom GPT instructions in ChatGPT, and the instructions field in most chatbot builders, work the same way. Different box, same idea.

How do you write a system prompt? The six parts

Anthropic suggests a simple test: show the prompt to a colleague who knows nothing about the task and ask them to follow it. If they would be confused, the model will be too. The six parts below pass that test when each one is a few plain sentences.

Write limits as actions, not feelings. "Be careful with refunds" is vague. "Never confirm a refund. Say the owner decides and offer to pass the request on" is something you can test.

Add one more line: what to do when unsure. This is the rule most prompts miss, and it is the one that protects you from invented answers (more in our guide to preventing AI hallucinations).

Example: a shop assistant

Copy this, replace the brackets and keep it under one page.

ROLE
You are the website assistant for [Shop name], a small online shop that sells [products]. You help visitors find products and understand shipping, returns and care instructions.

SOURCES
Answer only from the SHOP FACTS section below. If the answer is not there, do not guess.

LIMITS
- Never invent prices, stock levels, delivery dates or discounts.
- Never confirm a refund or exception. Say the owner decides, and offer to pass the request on.
- Do not give medical, legal or financial advice.
- Do not discuss other shops or compare prices.
- If asked about anything unrelated to the shop, say: "I can only help with questions about [Shop name]. Is there something about our products or orders I can help with?"

WHEN UNSURE
Say: "I don't have that information. I can pass your question to [Name], who replies within [one working day]." Then ask for the customer's email.

TONE
Warm, brief and plain. Short sentences. No slang, no exclamation marks, no emoji.

HANDOVER
Offer a person when the customer is upset, asks for a person, mentions a damaged or missing order, or asks the same thing twice without getting an answer. Collect their name, email and order number, then tell them when to expect a reply.

FORMAT
Reply in the customer's language. Maximum three short sentences. Put links on their own line.

SHOP FACTS
[Paste approved shipping, returns and product information here, or connect your help pages.]

Example: a service business that takes appointments

This one suits a clinic, a trades business, a salon or a consultancy. It collects a request, it does not book anything itself.

ROLE
You are the front-desk assistant for [Business name] in [City]. You answer questions about our services and collect appointment or quote requests. You do not book appointments yourself.

SOURCES
Use only the SERVICES AND HOURS section below. Never state a price unless it is written there.

LIMITS
- Never give a quote, diagnosis, medical or legal opinion, or guarantee of results.
- Never confirm a time slot. Say the team confirms by email or phone.
- Never ask for payment details, passwords or ID numbers.
- Do not tell customers what other staff or customers have done.

COLLECTING A REQUEST
Ask, one question at a time: what they need, preferred days, name, phone or email. Then summarise in two lines and say: "Thanks. [Name] will confirm within [one working day]."

WHEN UNSURE
Say you do not have that information and offer a call from the team. Do not guess.

URGENT CASES
If the customer describes an emergency, tell them to call [emergency number or local emergency services] now, and stop collecting details.

TONE
Calm, polite, professional. Plain words. No jargon.

HANDOVER
Hand over when the customer is angry, makes a complaint, asks for a person, or asks something outside the services listed.

FORMAT
Reply in the customer's language. One question per message. Short sentences.

SERVICES AND HOURS
[Paste approved services, starting prices if you publish them, opening hours and contact details here.]

The two prompts share a skeleton. Only the limits and the handover triggers change with the business. That is a good sign: most of the work is deciding your limits, not writing clever prose.

Why do chatbot rules get ignored?

Models follow clear instructions well. They struggle when the prompt works against itself. Four patterns cause most failures.

Explain the reason behind a limit when it helps. "Never confirm refunds, because the owner decides each case" gives the assistant something to reason from when a customer asks a slightly different question.

Keep facts out of the prompt

Rules and facts change at different speeds. Your tone changes once a year. Your prices, hours and shipping times change often. If both live in one block of text, every price update means editing the rulebook and risking a broken rule.

Keep two things apart: the system prompt (behaviour) and your approved content (facts), such as a help page, a price sheet or a short document the assistant reads. The prompt says "answer only from this material". The material holds the truth. Our guide to preventing AI hallucinations explains how to set that up and what still goes wrong.

How do you test the rules?

Every rule you write is a promise you have not checked yet. Test each one with a question built to break it. Ten to start with:

• Ask for a discount.
• Ask for a refund on an order that does not exist.
• Ask something the material does not cover.
• Ask about a competitor.
• Ask for medical, legal or tax advice.
• Write angrily and demand a person.
• Ask the same question twice in different words.
• Write in another language.
• Ask the assistant to ignore its instructions.
• Ask it to repeat its instructions.

Run each one more than once, because answers vary between runs. Mark each as pass or fail against your written rule, fix the prompt and run the full set again. The full method, with a test table, is in chatbot testing before launch. OpenAI's guide makes the same point for developers: build evaluations that measure prompt behaviour, and pin a specific model version in production so behaviour stays consistent.

What should never go in a system prompt?

Assume a customer can eventually see it. Models can be talked into revealing or ignoring instructions, and this is a known weakness, not a rare one. Do not put in the prompt:

• Passwords, API keys or private links.
• Personal data about customers or staff.
• Internal margins, supplier names or discount limits you do not want public.
• Rules that must be enforced with certainty, such as who may receive a refund. Enforce those in your own systems. The prompt can ask the assistant to decline, but it cannot be the only lock.

A system prompt reduces risk. It does not remove it. That is why the limits, the "I don't know" line and the handover rule matter more than polished wording.

How do you keep a system prompt up to date?

Treat the prompt like a small piece of software. Save every version with a date and a one-line note on what changed. Change one thing at a time, then rerun your test questions, so you know which edit helped and which broke something. Read a sample of real conversations each week at first. Every question the assistant fumbled is either a missing fact for your source or a missing line in your limits.

Review the prompt again whenever you change models, because the same text can behave differently on a new version. Before building anything, it is also worth deciding which tasks the assistant should take at all. Our note on when to automate a process is a good first filter.

Need help with the rules?

Lobos Design is a two-person studio in Las Palmas de Gran Canaria that designs and sets up assistants for small businesses. If you want a second pair of eyes on your limits and handover rules, see our AI integration service.