System Prompts

A prompt is a text input that determines how the model generates a response. There are two types of prompts:

• User prompt: the request text from the end user.
• System prompt: a text instruction that defines the agent’s behavior, including its goal, role, communication style, and limitations.

The system prompt, or instruction, affects the context in which the agent interprets requests. It is sent with every request to the model when using both the built-in widget and the API. A well-designed prompt improves the accuracy and relevance of responses.

System prompts allow you to adapt an agent for specific scenarios. For example:

• a domain expert;
• a specialized assistant;
• a character with a specific communication style.

Before deploying the agent in a project, you can test it in the playground and refine the instruction if necessary.

How to Add a System Prompt

The method for adding a system prompt depends on how the agent is used.

OpenAI-compatible API

When using the OpenAI-compatible API, the system prompt is provided with each request as a message with the system role.

Widget and Native API

For the widget and the native API, the system prompt is configured in the control panel. You can add an instruction:

• When creating the agent, by selecting one of the predefined examples or entering your own instruction;

• Later in the Playground section, in the settings of an existing agent.

Prompt Guidelines

To create the most effective instruction, include the following key elements:

• who the agent is (for example, a travel agent or a support specialist);
• its goals (answering questions, validating provided data, recommending services);
• what data it should rely on (specific resources or knowledge bases);
• its limitations (which topics it can discuss and which languages it can use).

1. Agent Identity

Define the agent’s identity by specifying its name, tasks, and area of expertise.

Good example:

Your name is Hosti. You are a virtual assistant that helps users find answers in the Hostman documentation and provides expert recommendations based on it.

Bad example:

You are a virtual assistant on the Hostman website.

2. Agent Goal

Define the goals as follows:

• specify the agent’s primary objective;
• list its key responsibilities and scope;
• describe its priorities, type of assistance, and limitations.

Good example:

Your main task is to help users understand and use Hostman services. You must provide accurate, clear, and complete answers based on the Hostman documentation.

Bad example:

Answer questions about Hostman services.

3. Agent Expertise

The agent’s expertise defines the areas and topics where it can provide accurate and relevant responses.

Use clear wording and:

• describe the agent’s specific capabilities;
• specify which data and knowledge bases it should rely on;
• teach the agent to provide examples when explaining complex topics, using verified sources and avoiding fabricated examples;
• instruct the agent to provide step-by-step instructions, formatting them with tables or code blocks when appropriate.

Good example:

You are a specialist in Hostman services and troubleshooting. Your primary focus is cloud databases, cloud servers, and Kubernetes.

Use the Hostman documentation as the primary source of truth and Hostman guides as a secondary source. Your knowledge should be based first on documentation and second on guide data.

Never invent information.

When answering technical questions, provide step-by-step instructions and use examples from the Hostman documentation or guides. Use tables when they help present information clearly.

Bad example:

You know everything about the Hostman platform based on its documentation. Do not invent answers. Add examples when helpful.

4. Topics and Communication Style

Define the scope of topics and the communication style:

• specify what the agent can talk about and which topics are restricted;
• indicate the primary language and supported languages;
• define the required tone and style of communication.

Good example:

Provide answers only based on Hostman documentation and guides. Do not discuss legal issues or handle complaints; in such cases, direct the user to support.

You understand and communicate in English, German, French, Spanish, Italian, Dutch, and Brazilian Portuguese. If a user contacts you in another language, politely ask them to rephrase the request in one of the supported languages.

Communicate politely and in a friendly manner.

Bad example:

Do not answer legal questions or complaints. Communicate only in English. Be polite.

5. Agent Limitations

Clearly define how the agent should behave if it cannot answer a question.

• The agent should acknowledge when it does not know the answer.
• The agent should ask clarifying questions if the request is unclear or incomplete.
• The agent should direct users to documentation or support when necessary.

Good example:

If you do not know the answer, say: “I don’t have enough information to answer this question. I’d be happy to help with something else.”

If the request is incomplete, ask for clarification: “Please provide more details about what you would like to do.”

If you are unsure about something, clearly state it and recommend consulting the Hostman documentation or contacting support.

Bad example:

If you cannot answer, inform the user and provide a link to the documentation.