Well-designed prompts and actions are the foundation of a reliable Agentic AI Bot. Follow these best practices to improve consistency, reduce hallucinations, and simplify maintenance.

System Prompts

A system prompt should be concise, clear, and focused. It defines the agent's behavior and is included with every user interaction.

Best Practices

• Use Markdown to structure instructions for better readability.
• Write positive instructions instead of negative ones.
  • Instead of: "Do not swear back to the user."
  • Use: "Always respond politely and empathetically."
• Clearly define:
  • The agent's role
  • Its tone, behavior, and personality
  • Any guidelines or restrictions for handling user queries
• Do not define actions in the system prompt. Actions are configured separately and made available to the agent automatically.
• Start simple. Add only the instructions required for the agent to perform its role. Introduce additional instructions only when necessary, and isolate them so they apply only to specific scenarios.
• Avoid business logic such as if-else conditions, case statements, or complex reasoning. Keep the prompt instructional rather than procedural.
• Keep it simple. The less reasoning the agent has to perform, the more consistently it responds.

Tip: Preview your Markdown using https://markdownlivepreview.com/.

Actions

Actions enable the agent to perform tasks such as calling APIs, retrieving data, or executing code. Well-defined actions help the agent choose the correct functionality.

Best Practices

• Use descriptive action names. The action name should clearly describe what the agent needs to do.
  • ✔ get_the_list_of_hospitals
  • ✘ list_hospitals
• Be consistent with terminology. Use either Action or Function throughout your implementation. Using Function consistently is recommended.
• Keep descriptions brief and clear. Focus on the purpose of the action and avoid unnecessary logic or implementation details.

Parameters

Well-defined parameters improve the accuracy of parameter extraction and reduce ambiguity.

Best Practices

• Use string as the parameter type for all parameters, even when another type may seem appropriate. This avoids platform-level limitations and ensures consistent parameter capture.
• If a short description isn't sufficient, refine it gradually while keeping it simple.
• If the agent struggles to capture basic values such as a name or email address, it is often a sign that the instructions are too complex. Simplifying the prompt usually improves parameter extraction.