Prompt Best Practices

1. Be Specific About the Agent's Role and Scope

The agent should know exactly what it's responsible for and what it's NOT responsible for. Avoid vague personas like "You are a helpful assistant." Instead, be concrete:

The more specific your persona, the more reliably the AI can fulfill its role.

2. Use Guardrails to Define What the Agent Should NOT Do

Guardrail prompts are powerful for preventing unwanted behavior. Instead of hoping the AI won't do something, explicitly tell it not to:

3. Use Schema Prompts to Describe Key Salesforce Fields

Schema prompts describe your Salesforce data structures. This reduces unnecessary tool calls (describe_object, get_picklist_values) that consume credits. Instead of the AI querying "What are the values of Opportunity.StageName?", you tell it upfront:

Use Schema prompts for the 10-20 most important fields and objects your agent references frequently.

4. Keep Each Prompt Focused on One Purpose

Don't try to cram multiple behaviors into one prompt. Create multiple focused prompts instead:

Focused prompts are easier to maintain, test, and debug.

5. Test Iteratively — Small Changes, Observe Behavior

Don't create 10 prompts at once and deploy. Instead:

This iterative approach helps you understand which prompts have the desired effect and catch unintended consequences early.

6. Use Context Variables for Personalization

Context variables like {USER_FIRST_NAME}, {TODAY_DATE}, and {USER_EMAIL} personalize interactions. Example:

This makes the agent feel responsive to the individual user, not generic.

7. Instruct for Concise Responses to Save Output Credits

Longer responses consume more output tokens, which increases cost. Instruct the AI to be brief:

Brief, focused responses are often better for users anyway — they reduce cognitive load.

8. Avoid Repeating Information Across Multiple Prompts

Redundant instructions consume extra tokens and can lead to confusion. If you've already stated something in one prompt, don't repeat it in another. Cross-reference if needed:

9. Use Sort Order to Control the Order Prompts Are Assembled

If some instructions should take priority over others, use lower Sort_Order numbers. Example:

• Sort_Order 10: Critical guardrails (safety rules) — processed first
• Sort_Order 20: Schema information (data context)
• Sort_Order 30: Custom instructions (behavioral rules)
• Sort_Order 40: Action instructions (specific actions)

Later prompts can build on earlier ones. Arrange them logically.

10. Document Your Prompt Strategy

Create internal documentation explaining:

• What each prompt does and why it exists
• The intended user base for the agent
• When to activate/deactivate specific prompts
• How prompts work together to shape behavior
• Any known limitations or edge cases

This helps other admins understand your design decisions and maintain the agent effectively.

Advanced Techniques

A/B Testing Prompts

Test two versions of a prompt by:

1. Create Prompt A (active)
2. Create Prompt B (inactive)
3. Run both through the agent for a week each, tracking results
4. Decide which performs better and keep the winner

Progressive Refinement

Start with a minimal persona and gradually add prompts as you learn what works:

1. Week 1: Persona + basic Custom Instructions
2. Week 2: Add Schema prompts based on queries the AI makes
3. Week 3: Add Guardrails based on unwanted behaviors observed
4. Week 4+: Fine-tune and optimize

Context-Specific Overrides

Create Persona_Override prompts for specific contexts. For example, a Sales agent might have:

• Base Persona: General sales assistant tone
• Persona_Override (Sort 10): More formal tone for Enterprise customers
• Persona_Override (Sort 20): More casual tone for SMB customers

The AI can learn to apply the appropriate tone based on context.

Frequently Asked Questions

Next Steps

Explore custom prompts per agent to organize your prompts, or learn about context variables for advanced personalization.