Skip to main content
The prompt is the core of Agent. A well-written prompt can make a agent reliable and effective, while a poorly written one can break the entire agent. This guide shows how to use structured prompt engineering to build agent that behave consistently and perform tasks correctly.
To see examples, explore prompts used in Agent templates.

​
Structure Your Prompt

Avoid putting all instructions into a single block of text.
Using a modular structure helps the AI model clearly understand its role, scope, and boundaries. We recommend organizing every agent prompt into three core sections:
  • Role: Define what responsibility the agent holds and how the agent behaves.
  • Tasks: Describe the step-by-step workflow the skill should follow to complete its job.
  • Rules: Set constraints, priorities, and edge-case handling to prevent unwanted behavior.
# Role
You are a assistant for... You are responsible for handling...
Your style is polite, warm...

# Tasks
## Task1:...
1.
2.
## Task2:...
1.

# Rules
1. You should...
2. Never...

​
Write Tasks as a Process

When writing the Tasks section, think like a developer designing logic. Break execution into clear steps and use structured flows. Numbered steps and conditional logic (If / Else) significantly improve reasoning accuracy and execution stability. Task example: 
# Tasks

## Task1: Appointment Scheduling

1. Extract user’s name and desired appointment time slot from the message.

2. Ask for Missing Information(If Needed)
- If the user has provided both name and desired time slot, skip this step and proceed to Step 3.
- If the user has provided only one piece of information or none, politely ask the user to supplement the missing details.

3. Check Availability via google_calendar
- If user’s slot is within available time, Proceed to Step 4.
- If conflict, Reply exactly: "The selected time slot is unavailable due to a conflict. Please confirm a new time slot."
  - Re-run Step 3 if user provides a new time slot.

4. Call google_calendar to add the confirmed time slot linked to the user’s name, then send a booking confirmation.

​
Recommendations

  • Be Specific and Detailed
    Avoid vague instructions. Clearly define context, expected output, tone, format, and length.
    ❌✅
    Ask the user for appointment time.Ask the user for 15-min dental checkup time slot (Mon-Fri 9AM–5PM, 30-min intervals only) in 1 polite sentence.
  • Use Positive Guidance
    Instead of telling the agent what not to do, tell it what it should do.
    This keeps the model focused on the desired behavior.
    ❌✅
    Don’t ask for irrelevant info.Ask only for guest name, check-in date, and desired spa time slot.
  • Provide Few-Shot Examples
    Include 1–2 complete examples of ideal user–agent interactions.
    This helps the model learn tone, structure, and response length through imitation.
  • Emphasize Critical Instructions
    For extremely important rules or steps, use UPPERCASE to draw attention.
  • Write in English When Possible
    While Solvea supports multiple languages, English prompts currently achieve the highest instruction-following accuracy and reasoning quality.

​
Add Tools and Knowledge in the Prompt

You can reference tools, knowledge sources, and handoff actions directly in the prompt using / . This allows the agent to:
  • Retrieve knowledge
  • Handoff to a human agent. (reason and action are required)
  • Call tools
Prompt Example： 
# Retrieve knowledge
- call retrieve_knowledge tool to answer customer's inquiries.

# Handoff to a human agent
- When the customer still cannot provide order number ,pleasse call Only Handoff.

# Call Tools
- use all of the tracking numbers to call logistics_inquiry tool to get the logistics details.
- use the order number to call shopify tool to get the tracking numbers of the order
- call google_calendar tool to add the event.
- call google_calendar tool to check if the time slot is available
- call google_sheet tool to retrieve the file "xx" and add a new row to the sheet.