To get good results from Agentforce, the most important skill is prompt engineering.
In this post, you will learn practical prompt engineering tips that help you design prompts that are clear, reliable, safe, and easy to maintain.
1. Start With a Clear Goal
Every prompt should have one clear purpose.
Before writing your prompt, ask yourself:
- What problem should the agent solve?
- Who is the user?
- What is the exact output needed?
Example
Goal: Create a support case.
Prompt:
'Create a new support case for the customer. Include name, issue summary, and priority.'
Clear goals reduce confusion and produce consistent results.
2. Give Proper Role and Context to the Agent
Agents respond better when you tell them who they are and what they should do.
Example
'You are a Salesforce Support Agent. Your task is to help users create and update support cases in the system.'
This sets the direction and avoids unwanted steps.
3. Add Rules and Constraints
Rules help the agent operate in a controlled and predictable way.
Examples of rules
- Do not guess data.
- Only use information received from the user or Salesforce records.
- Ask questions when inputs are missing.
- Follow data security and privacy requirements.
Clear constraints prevent incorrect actions.
4. Provide Examples for Better Outputs
Examples improve accuracy.
They show the format, tone, and structure you want.
Example
Input: 'Customer cannot access their account.'
Output format:
- Issue Type: Login Issue
- Priority: Medium
- Action: Create a support case
You can provide multiple examples depending on complexity.
5. Use Step-by-Step Instructions
Breaking a task into steps helps the model think in an organized way.
Example
-
Read the user request.
-
Extract relevant details.
-
Validate the details.
-
Ask follow-up questions if needed.
-
Provide the final action or answer.
This method reduces errors and produces consistent results.
6. Tell the Agent What Not To Do
Negative instructions are important.
They help reduce hallucinations and protect data usage.
Examples
- Do not create records without confirmation.
- Do not include personal or sensitive data unless explicitly provided.
- Do not make assumptions.
This aligns with Salesforce’s AI Trust Layer principles.
7. Keep Prompts Short and Focused
Long prompts are harder to maintain and slow down performance.
Use simple language and break large tasks into smaller prompts or sub-prompts.
A good prompt is:
- Short
- Specific
- Easy to read
- Easy to update
8. Use Slot-Based Prompts for Structured Inputs
When you expect specific fields, clearly list them.
Example
'Provide the result in this format:
- Name:
- Email:
- Issue:
- Priority:'
This prevents the agent from generating unwanted content.
9. Test Prompts Across Different Scenarios
Do not rely on single-case testing.
Test for:
- Missing information
- Wrong information
- Long messages
- Short messages
- Edge cases
- Multi-step interactions
Testing helps you refine and improve prompts before deployment.
10. Keep Improving Your Prompt Over Time
Prompt engineering is iterative.
Review agent logs to check:
- Incorrect answers
- Repeated questions
- Confusing outputs
- Missing context
Update your prompts based on real user interactions.
Bonus: Salesforce-Specific Best Practices
1. Use the Salesforce Trust Layer by Design
Prompts should always assume that safety, grounding, and policy checks are happening by default.
Write prompts that encourage grounded answers based on approved data.
2. Reference Salesforce Objects and Fields Clearly
Instead of saying:
'Get customer details'
Say:
'Fetch the Contact record using the email provided by the user.'
3. Plan for Multi-Turn Conversations
Agentforce agents often handle multi-step tasks.
Write prompts that allow the agent to ask intelligent follow-up questions.
Watch Complete Video Below
0 Comments