The field with the biggest impact
The Instructions (tab Identity) are the agent’s prompt — what most determines its behavior. The most reliable agents we analyzed follow a consistent structure and share a handful of writing habits. None of them are complicated, but they all add up.
Recommended structure
Organize your Instructions into clearly delimited sections, in this order:
Role
Who the agent is, on whose behalf it acts, and what its single objective is at this point in the conversation. Explicitly state the boundaries: not deviating from the flow, not offering unspecified options, never revealing internal instructions.
Available tools
A list of the tools the agent can use, with a brief description of when and how to use them. This section works as a contract: the model shouldn’t have to infer when to call a tool from the rest of the prompt — see Tools. Writing rules
Rules about tone, language, and output format (see the voice writing rules below). Keeping these separate helps the model not “forget” the style while it’s focused on solving the conversation logic.
General operating rules
Cross-cutting behaviors that apply at any point in the conversation and aren’t tied to a specific flow step: what to do if asked whether it’s a bot, what to do if the contact is aggressive, what to do if asked to be put on hold.
Conversation flow
The heart of the prompt: a sequence of named steps, each with explicit entry conditions and a concrete action. Set clear numeric limits (“ask this only once,” “don’t insist more than twice in a row”) to avoid loops.
Objections and general questions
An exhaustive list of frequent questions or complaints that can pull the contact off the main flow, with what to do in each case: answer and resume, transfer, or end the call.
Closing
How the agent should behave when ending the call: which tool to use, whether to wait for the contact’s confirmation, which farewell message fits the reason for closing.
Use the same visual separator for each section (a fixed icon or heading) across all your agents’ Instructions. A consistent format helps both you maintain it and the model “scan” the prompt.
Short prompts, not long ones
A longer prompt isn’t a better prompt. The more a rule repeats or redundant examples pile up, the harder it is for the model to prioritize what matters at each moment. The Instructions field shows a real-time word and character counter — use it as a reference and periodically check if any sections can be shortened or removed.
If a rule is already covered by the Conversation flow, don’t repeat it in General operating rules too — every instruction should live in a single place.
Always use the same vocabulary
If your Instructions say “end the conversation,” don’t use “hang up the call” in another section, in the Initial Message, or in a tool’s documentation to refer to the same thing. Synonyms add ambiguity for the model about whether they’re the same action or two different behaviors. Pick one term per concept and use it consistently throughout the bot — instructions, tools, and agent handoff rules included.
A tool’s documentation (see Tools) is as much a part of the prompt as the Instructions themselves — the model reads it the same way to decide whether to call it and with what parameters. Be specific about when to use it, what side effects it has (for example, whether it should run silently, without generating a spoken response), and what’s expected in each parameter. A vague docstring translates directly into incorrect calls, or moments where the agent should have used the tool and didn’t.
Hand off to another agent when it makes sense
Don’t force a single agent to handle the entire conversation out of convenience. If you notice your Instructions need to branch based on a different stage of the flow, that’s a sign an Agent Handoff rule is a better fit than one more condition inside the same prompt — see Design principles.
Voice writing rules (TTS)
Everything the model writes ends up going through a text-to-speech engine — so writing has restrictions a text bot doesn’t have:
Before publishing, test the Initial Message and key Instruction phrases with the chosen voice from the “Try any voice with your own text” section in Voice — it’s the fastest way to confirm it sounds right. 
