📖 Glossary

What is a system prompt?

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

In short

A system prompt is the written instructions given to an AI agent that shape its role, tone, knowledge boundaries and rules before any customer conversation begins. It works like a job description and a training manual combined, telling the AI who it represents, what it should and should not answer, and how to behave when unsure.

Key takeaways

A system prompt is set once per agent and applies to every conversation, unlike a normal chat message.
Good ones cover role, tone, scope, what to do when unsure, and escalation rules.
Vague system prompts are the single biggest cause of an AI agent giving wrong or off-brand answers.
Prompts should be versioned and tested after every change, not edited live against real customers.
A system prompt is not the same as a knowledge base — it is instructions, not reference documents.

The definition

A system prompt is a block of instructions given to an AI model that stays in effect for the entire conversation, set by the business rather than the customer. It is distinct from the messages the customer types or says, which the AI treats as input to respond to, not as instructions to follow. Think of it as the job brief you would hand a new employee on day one: who they are, what they are allowed to do, and how to act when a situation is not covered. Unlike a human employee, the AI cannot ask a colleague for clarification mid-conversation, so an unclear prompt tends to surface as an unclear or wrong answer immediately, rather than a question back to you first.

Every reply the AI gives is shaped by this prompt, even though the customer never sees it directly.

It sits underneath every single conversation the agent has, quietly, unlike a normal message which is sent once and then gone. Get it right once and the improvement shows up across every future customer interaction; get it wrong and the same mistake repeats at scale.

The sections a good system prompt covers

Role and identity: who the AI is ("You are Riya, the WhatsApp assistant for Sharma Opticians").
Tone: formal, friendly, concise; whether to use emojis; which language to default to.
Scope: exactly what topics it should answer (store hours, product details, appointment booking) and what it must not (medical advice, discounts outside policy).
Behaviour when unsure: say so honestly and offer to connect a human, rather than guessing.
Escalation rules: specific triggers, like an angry customer or a refund request, that should hand off to a person.

Not every prompt needs all five in equal depth, but skipping scope and escalation rules is where most businesses run into trouble, because the AI has nothing telling it where to stop.

A worked example

A short but usable system prompt for a salon's WhatsApp bot might read:

"You are Meera, the booking assistant for Glow Salon in Andheri. Reply in a warm, brief tone, in the same language the customer uses. Help with service pricing, availability and booking appointments. If asked about medical skin conditions, say this needs an in-person consultation with our dermatologist. If a customer is upset or asks for a refund, say you are connecting them to the team and stop responding."

Notice it defines identity, tone, scope, a boundary, and an escalation trigger, all in a few sentences.

You do not need every prompt to read like a legal document. Plain, direct sentences work better than jargon, because the AI is reading and following them the same way a new employee would read a briefing note on their first day.

Common mistakes

Too vague: "Be helpful and answer questions" gives the AI no real boundaries, so it may answer things it should not.
No escalation rule: without one, the AI keeps trying to help even when a human is clearly needed.
Contradicting the knowledge base: instructions that conflict with uploaded documents confuse the AI's answers.
Never updated: a prompt written once at launch and never revisited drifts out of date as pricing or policy changes.
Editing live: changing the prompt directly on the production agent without testing risks breaking answers customers are relying on right now.

Another quiet mistake is writing the prompt once at launch and treating it as finished. Pricing changes, new products launch, and policies get updated. If the prompt is not part of that same update process, the AI keeps confidently repeating outdated information.

Versioning and testing

Treat the system prompt like any other important business document: keep a record of what changed and why, test changes before they go live, and roll back quickly if a change makes answers worse. This matters more as an agent's scope grows: a small wording change can shift behaviour across hundreds of future conversations at once.

A simple habit that helps: keep a short changelog, even a plain note of "changed refund policy wording on 12 Sept", alongside the prompt itself. When an answer looks wrong weeks later, that record makes it far faster to work out whether a recent change caused it.

System prompts on Convo360

On Convo360, every AI agent has its own system prompt field, kept separate from its knowledge base documents and its temperature setting. Convo360 keeps drafts vs published versions with roll-back, so you can edit and test a new prompt using Preview / Try-it (a dry run) before it reaches real customers, and revert instantly if something goes wrong.

The system prompt also works alongside temperature, a per-agent setting that controls how strictly the AI sticks to predictable phrasing versus how much variation it allows in its wording, useful to know when two agents with the same prompt still sound slightly different from each other in everyday conversation, particularly on longer, more open-ended replies.

Example conversation

Can you give me a bigger discount than what's listed?

I'm not able to offer discounts beyond our published pricing, but I can connect you with our sales team who may have more flexibility. Would you like me to do that?

Write your first AI agent prompt

Start free in minutes — no code, no credit card. Bring your own AI keys and pay no per-message markup.

Start 14-day free trial →See pricing

Frequently asked questions

Is a system prompt the same as a script?

Not quite. A script dictates exact words for exact situations. A system prompt sets rules, tone and boundaries, and the AI generates its own wording within those rules for whatever the customer actually says.

How long should a system prompt be?

Long enough to cover role, tone, scope and escalation clearly, but no longer than needed. Overly long prompts with repeated or conflicting instructions can confuse the AI rather than helping it.

Can I have different system prompts for different channels?

Yes, and it is often useful — a WhatsApp agent might be more casual in tone than a voice agent, even if both represent the same business and share the same knowledge base.

What happens if I don't write a system prompt at all?

Most platforms apply a generic default, which rarely matches your business tone or boundaries well. Writing even a short, specific prompt significantly improves answer quality and consistency.

Do customers ever see the system prompt?

No. It is instructions for the AI, not a message shown to customers. Only the AI's actual replies, shaped by those instructions, are visible in the conversation.