Skip to main content
This guide covers the most important advice and the first things to check when improving a single prompt agent’s behavior. It is not an exhaustive guide. Start with the general prompting principles, which apply to both the prompt and the tools you give the agent. The prompt should describe the whole task while allowing the agent to adapt to how the caller responds. Give it the job to accomplish, the information it needs, and the requirements it must follow. Single prompt agents are flexible by default and can handle common situations such as callers changing their minds or answering questions out of order. You generally don’t need to write separate instructions for each of these situations. If you are deciding which agent type to build, see When to use single prompt vs. conversation flow agents.

​
Only prompt for behavior the LLM controls

The prompt primarily guides what the agent says and when it calls tools. Use settings to control speaking tone, background-noise filtering, and interruption handling. For example, “Keep speaking even if the caller talks over you” will not change interruption handling; adjust the interruption settings instead. There are specific exceptions for speech delivery. With Expressive Mode enabled on a platform voice, you can prompt the agent to output emotion and effect tags such as [happy] and [pause]. You can also prompt the agent to use spaced dashes to add pauses or read slowly. These supported controls let prompts affect delivery, but do not replace the agent’s audio settings.

​
Choose and describe tools

Keep the number of tools small and their definitions concise. Tool definitions are part of the model’s context, so the general prompting principles apply to tool names, descriptions, and parameter descriptions too. Include only the tools needed for the task and avoid unnecessary detail or repetition. This reduces the amount of context and the number of choices the model must work through, helping it select tools and follow instructions more reliably. When deciding where to put tool instructions, we recommend describing what the tool does in its description and when to call it in the general prompt. This keeps tool-use instructions alongside the task instructions they belong to. See function calling for tool configuration and basic settings for the agent’s model, voice, and call settings.