Skip to main content

Prompt Engineering

The quality of your AI agent depends heavily on the system prompt. This guide covers best practices for writing effective prompts.

System Prompt Structure

A well-structured system prompt has four sections:

1. ROLE — Who the agent is
2. CAPABILITIES — What it can do
3. CONSTRAINTS — What it should NOT do
4. FORMAT — How to structure responses

Example: E-Commerce Support Agent

ROLE:
You are a customer support agent for an online store. You help customers with order tracking, returns, and product questions.

CAPABILITIES:
- Look up orders using the lookup_order tool
- Process return requests using the create_return tool
- Answer product questions from the knowledge base
- Escalate issues by creating support tickets

CONSTRAINTS:
- Never share other customers' information
- Do not process refunds over $500 without escalation
- Do not make promises about delivery dates you cannot verify
- If unsure, say "Let me connect you with a specialist" rather than guessing

FORMAT:
- Keep responses under 150 words
- Use bullet points for lists
- Include order IDs and tracking numbers when available
- End with "Is there anything else I can help with?"

Variable Injection

Use variables in your system prompts to customize behavior dynamically:

You are a support agent for {{company_name}}.
The customer's name is {{customer_name}}.
Their account type is {{account_type}}.
Today's date is {{current_date}}.

Variables are replaced at runtime with actual values from your workflow data.

Temperature Settings

TemperatureBehaviorUse Case
0.0Deterministic, consistentData extraction, classification, factual Q&A
0.3Slightly varied, mostly consistentCustomer support, technical writing
0.7Creative with good coherenceBlog posts, marketing copy
1.0Highly creative, less predictableBrainstorming, creative writing

Tips

  • Be specific — "Respond in 2-3 sentences" beats "Be concise"
  • Give examples — Show the agent what good responses look like
  • Define boundaries — Explicitly state what the agent should refuse
  • Test edge cases — Try ambiguous inputs, empty inputs, and adversarial prompts
  • Iterate — Prompt engineering is iterative. Test, review, and refine.