Hi all, can someone please help me improve my instructions/prompts to the agent. Here is a link to the agent instructions/prompt I'm using now https://docs.google.com/document/d/1UL7qWLXW-RrPKP0FoQCfaWpmNGw5AArUr8ZiLdeSjvE/edit?usp=sharing. The

{}

you see in my instructions is because I'm using context variables in my agent. For context, I'm creating an outbound agent that calls interested leads, talks through their problem, and books an appointment for them. I'm using the chained architecture specifically Cartesia for the STT(ink-whisper ) & TTS(sonic-3) and Gemini (gemini-3-flash-preview) as the LLM. Using a calendar MCP to book the appointment. In my testing, I have noticed the following from my agent; 1. After the initial greeting, the agent directly goes to ask questions without having a good conversation about the issue/problem the user filled in the form. 2. The agent asks multiple questions at once without waiting for the user to answer. 3. The agent keeps quiet when executing the MCP tools like creating event on the calendar. Check this agent recordings to see how my agent behaves; • https://storage.googleapis.com/voice-ai-call-recordings/new-room/agent-2026-02-24T152721-EAWUWy3o9Dpz.ogg • https://storage.googleapis.com/voice-ai-call-recordings/new-room/agent-2026-02-24T151037-UeynTurzslzb.ogg How I want the agent to behave; 1. Greet the user. Use the context variables for a customized conversation 2. Discuss the problem with the user. Have a natural empathetic conversation 3. When taking action on behalf of the user, inform the user what you're going to do. Ask the user for their confirmation. 4. When the agent is asking the user questions, I want the agent to ask a question, wait for the answer from the user and then proceed to the next question. 5. When executing a tool with the MCP like, I want the agent to be talking or inform the user to wait as the agent takes an action. I don't it to just be quiet. I believe if I improve the instructions, the agent behaviour will also improve. Please help to modify the instructions. Yes, I followed the prompting guide when creating these instructions. I have provided you all comment permission on the above Google Doc, please suggest how I can improve the agent instructions.

I can help refine your agent prompt so it follows a more natural, empathetic conversation flow and properly narrates actions during MCP tool execution. Based on your observations, the main issue is that the instructions likely don’t strongly enforce turn-taking, confirmation before actions, and verbal status updates while tools run. I’ll review your Google Doc and rewrite the instructions into a tighter system prompt that guides the agent to greet with context variables, explore the lead’s problem conversationally before asking structured questions, wait for each response before continuing, and clearly inform the user when it’s booking or performing any calendar action. @fresh-hospital-14455