Writing a system prompt for a business bot: tone, rules and guardrails

Updated 2026-09-19 · 4 min read · By the Convo360 team

Prerequisites and the five-part structure that works

You need an AI agent already created (choose OpenAI, Claude or Gemini as its brain; see our model comparison if you haven't decided). It also helps to write down, in plain language, the three or four things this specific bot must never do, before you start drafting formal prompt sections, since those are usually the rules that matter most.

A system prompt is the set of instructions an AI agent reads before every conversation. Vague prompts produce vague, inconsistent bots. A structure that consistently works well:

1. Role: who the bot is and what it's for, in one or two sentences.
2. Tone: how it should sound: formal, warm, brief, chatty.
3. Do/don't list: specific, concrete rules about what it can and cannot say or do.
4. Language rules: which languages to use, and how to handle Hinglish or regional-language messages.
5. Handoff rule: exactly when to bring in a human instead of answering itself.

Role and tone, and do/don't lists where guardrails live

Instead of 'You are a helpful assistant', write something like: 'You are the WhatsApp assistant for Sunrise Dental Clinic, Andheri. You help patients book appointments, answer questions about services and clinic timings, and never give medical advice.'

For tone, name a reference point: 'Warm and reassuring, like a friendly front-desk receptionist, not overly formal, but always respectful. Keep replies under 3 sentences unless explaining something detailed.'

The do/don't list is the section that prevents embarrassing or risky replies. Be concrete:

• Do: Always confirm appointment date and time back to the patient before ending the booking.
• Do: Offer the clinic's phone number for anything urgent or an emergency.
• Don't: Never give medical advice or diagnose symptoms. Always suggest booking a consultation instead.
• Don't: Never ask for or accept OTPs, full card numbers or passwords over chat.
• Don't: Never promise a specific treatment outcome.

Language rules and the handoff rule

State languages explicitly: 'Respond in the same language the customer writes in: Hindi, English or Hinglish. If a customer writes in another Indian language you're less confident in, reply in simple English and ask if they'd prefer that language.'

And always include a clear handoff rule: 'If a customer asks something outside clinic services, seems frustrated, or explicitly asks for a person, respond: "Let me connect you with our team" and flag the conversation for live-chat handoff.' Convo360's live-chat inbox pauses the bot automatically the moment a human team member replies, so this handoff is seamless.

A full worked example: clinic system prompt

'You are the WhatsApp assistant for Sunrise Dental Clinic, Andheri, Mumbai. You help patients book, reschedule or cancel appointments, and answer questions about services, timings and location. Tone: warm and reassuring, like a friendly front-desk receptionist. Keep replies under 3 sentences unless explaining a process.

Do: confirm date and time back to the patient; offer the clinic's phone number for emergencies. Don't: give medical advice, diagnose symptoms, discuss pricing beyond the listed consultation fee, or ask for OTPs or payment details.

Respond in the language the patient writes in (Hindi, English or Hinglish). If they seem frustrated, confused, or ask for a person directly, say "Let me connect you with our front desk team" and hand off.'

Exact click path: setting and testing it on Convo360

1. Open your AI agent settings and paste the structured prompt into the system prompt field.
2. Attach a knowledge base if the bot should answer from documents, and reference it in the prompt: 'Use the knowledge base for pricing and service details.'
3. Set an appropriate temperature: lower for a precise, guardrail-heavy bot, higher for a more conversational one.
4. Use Preview / Try-it to test tricky edge cases: an angry customer, an off-topic question, a request for medical advice, and refine the prompt based on what breaks.

Troubleshooting common issues

What to test before publishing

Run at least five deliberately tricky questions through Preview / Try-it before publishing: an angry or frustrated message, a question completely outside the bot's scope, a direct request for something on the 'don't' list (like medical advice or a discount), a message in a language you're unsure about, and an explicit 'let me talk to a person' request. If any of these produce a response that breaks a rule, tighten that specific section of the prompt and retest. This five-minute check catches most embarrassing failures before a real customer ever sees them.

Step by step

1. Write the role sectionDescribe who the bot is and what it's for in one or two specific sentences.
2. Define the toneName a concrete reference point and a reply-length guide.
3. Write the do/don't listList specific, concrete rules rather than vague guidance.
4. Add language rulesState which languages to use and how to match the customer's language.
5. Add the handoff ruleSpecify exactly when the bot should bring in a human, with trigger phrases.
6. Paste it into the agent settingsAdd the full prompt to your AI agent's system prompt field.
7. Test edge casesUse Preview / Try-it with tricky, off-script questions before publishing.

Example conversation

Frequently asked questions

Related reading

Available across India — see WhatsApp & AI calling automation by state and city.