Documentation / Prompting guide
Prompting guide
Define the goal, tone, boundaries and tool usage for consistent conversations.
Instructions tell the agent how to conduct a conversation. Start with one specific goal and test real situations before adding more rules.
Before you start
Create an agent and open its instructions. The currently available agent type uses a prompt: write its goal, tone, conversation steps and rules in the agent instructions.
Structure your instructions
Separate role, goal, conversation, tools and boundaries. Ask one question at a time and explain what to do when information is missing.
# Role
You support customers of a store using short, clear sentences.
# Goal
Help customers check an order's status.
# Conversation
Ask for the order number if it has not been provided.
Confirm the number before looking it up. Ask one question at a time.
# Tools
Use lookup_order only after obtaining the order number.
Explain the result based on the tool response.
If the lookup fails, say you could not verify the status.
# Boundaries
Do not invent delivery dates, status or update confirmations.
If you cannot resolve the issue, explain how to contact the team.
A tool instruction requires a tool that exists and is attached to the relevant step. Configure its name, description and parameters in HTTP API or MCP.
Personalize with context
Verify the behavior
Test a valid order, a missing number, a failed lookup and an unrelated question. Check what the agent said and the actions it performed. Change one rule at a time.
Silence timing, duration and interruptions are conversation settings. Tone instructions do not replace these controls. Prompt boundaries do not replace authorization and validation in your external system.
Continue with Testing and Publishing.
