Writing an Effective Custom Agent System Prompt

What a System Prompt Does

The system prompt controls how a custom agent behaves.

It defines:

• The role of the agent

• The type of tasks it should handle

• The rules it should follow

• The tone and style of responses

Unlike user messages, the system prompt is always applied and shapes every response the agent produces.

Start with a Clear Role

Every system prompt should clearly define the agent’s role.

A good role description:

• Is specific and task-focused

• Matches the agent’s scope

• Avoids generic wording

Good examples:

• “You are a customer support agent answering product-related questions.”

• “You are an internal assistant that summarizes company documents.”

Avoid vague roles like “helpful assistant” or “general AI”.

Describe the Agent’s Responsibilities

After defining the role, clearly state what the agent is responsible for.

Good responsibility definitions:

• Focus on one type of task

• Match how users are expected to use the agent

• Do not overlap with unrelated use cases

Clear responsibilities help reduce inconsistent or irrelevant answers.

Add Rules and Constraints

Rules and constraints help keep agent behaviour predictable.

Use them to define:

• What the agent should not do

• How to handle missing information

• When to ask for clarification

• When to refuse or redirect a request

Constraints are especially important for keeping agents within their intended scope.

Define Tone and Output Style

Tone and output style should be defined explicitly.

For example:

• Professional or casual tone

• Short or detailed responses

• Use of bullet points or paragraphs

Defining style helps ensure consistent output across different users and requests.

Keep the Prompt Clear and Focused

Effective system prompts are clear and concise.

Avoid:

• Overly long instructions

• Contradictory rules

• Trying to handle multiple responsibilities in one prompt

If a prompt becomes complex, it’s often a sign that the agent’s scope should be revisited.

What to Read Next

• Tools, Permissions, and Knowledge Access

• Defining Custom Agent Scope

• When to Create a Custom Agent (and When Not To)