Can website visitors see the system prompt?

Not directly. The system prompt is sent to the model by the server and is not shown in the chat. Users sometimes manage to make a model repeat parts of it, so treat it as private but not secret.

How long should a system prompt be?

Long enough to cover role, scope, tone and key rules, and short enough that every line matters. For a website chatbot, a few short paragraphs or a bullet list is usually enough. Detailed product facts belong in the knowledge base rather than the system prompt.

What is the difference between a system prompt and a user prompt?

The system prompt is written by whoever runs the chatbot and applies to every conversation. A user prompt is what the person types in a single message. Models are generally trained to follow the system prompt when the two conflict.

Should I put my product information in the system prompt?

Only short, stable facts. Detailed or changing information such as product lists, prices and policies belongs in the knowledge base, where retrieval finds the relevant part for each question. That keeps the system prompt focused and easier to maintain.

System Prompt

Last updated: October 6, 2026

What a system prompt is

Chat-based language models separate messages by role. The system message, or system prompt, comes from the developer or the business running the chatbot. User messages come from the person chatting, and assistant messages are the model’s own replies. Most models are trained to give the system prompt more weight than user messages, which makes it the main place to configure behavior.

Products often give it a friendlier name, such as instructions, persona or custom instructions.

What a system prompt usually contains

A good system prompt for a business chatbot covers a few areas:

Many applications also add text automatically, such as retrieved documents, tool descriptions or built-in rules, so the final prompt the model sees is longer than what the owner wrote.

Role: who the assistant is and whom it serves.
Scope: what it should help with and what is out of bounds.
Tone and format: formal or casual, short or detailed, lists or prose.
Rules: how to handle missing information, which topics to avoid and when to suggest contacting the team.
Stable facts: details that rarely change, such as the support email address.

Example

A garden center could use this system prompt: “You are the assistant of Casa Verde, a garden center. Help visitors with questions about plants, opening hours and delivery, using only the provided content. Keep answers short and friendly. If you are not sure, say so and suggest calling the store. Do not give advice on pesticides beyond what the product pages say.” With these instructions, a question about a pest problem gets an answer based on the shop’s own product information, and an unrelated request, such as help with homework, is politely declined.

Why the system prompt matters for business chatbots

The system prompt is the simplest way to change how a chatbot behaves. A clear one keeps the bot on topic, gives it a consistent voice and tells it what to do when it does not know.

It is not a security boundary, though. Determined users can sometimes extract or bypass instructions through prompt injection, so a system prompt should never contain passwords, API keys or confidential information. Rules that really matter need protection outside the prompt as well.

System prompts in intoCHAT

In intoCHAT, the system prompt is the Instructions field of your agent, and you can start from templates. intoCHAT combines your instructions with runtime settings, such as available actions and the lead-capture mode, and with built-in rules on grounding, privacy and prompt-injection resistance. The built-in rules take precedence where they conflict with your text; otherwise your role, tone, language and scope are kept. The welcome message and suggested questions are set separately, and you can test changes in the playground before they go live.