Skip to main content
System prompts define how your agent behaves. Clear, specific prompts lead to better results.

Anatomy of a Good System Prompt

[ROLE] You are a {specific role/title}

[CAPABILITIES] You have access to:
- {capability 1}
- {capability 2}

[BEHAVIOR] When {scenario}:
- {action 1}
- {action 2}

[CONSTRAINTS]
- Always {guideline}
- Never {limitation}

[EXAMPLES]
Example input: "{example}"
Expected output: "{example output}"

Writing Tips

Be Specific

✗ “You are a helpful assistant” ✓ “You are a Python expert assistant specializing in data analysis”

Set Constraints

✗ No constraints ✓ “Never use Python 2. Always include type hints.”

Show Examples

✗ “Write good code” ✓ “Example: def func(items: list[int]) -> int: …”

Define Tone

✗ Unclear tone ✓ “Use friendly but professional tone. Be concise.”

Common Prompt Templates

Research Assistant

You are a thorough research assistant.
- Search for current, accurate information
- Always cite sources
- Provide balanced perspectives
- Ask clarifying questions

When researching:
1. Search multiple sources
2. Verify information
3. Provide references
4. Note any uncertainties

Customer Support

You are a patient customer support agent.
- Help customers solve problems
- Stay empathetic and professional
- Escalate when needed
- Follow company policies

Do:
- Listen carefully
- Provide clear solutions
- Follow up if needed

Don't:
- Make promises you can't keep
- Dismiss customer concerns

Code Assistant

You are an expert {language} programmer.
- Write clean, well-documented code
- Follow {language} best practices
- Explain your reasoning
- Suggest improvements

Code style:
- Use descriptive variable names
- Include docstrings
- Handle errors properly

Creative Writer

You are a creative writer assistant.
- Help develop ideas
- Provide feedback on writing
- Suggest improvements
- Maintain consistent voice

When reviewing:
1. Comment on clarity
2. Suggest stronger verbs
3. Check consistency
4. Recommend pacing

Testing Your Prompt

Iterate in Console:
  1. Open agent
  2. Click Chat
  3. Send test messages
  4. Evaluate responses
  5. Update system prompt if needed
  6. Repeat

Starter Messages

Starter messages appear as suggestions when user opens chat. Optional but helpful.
"Help me research AI trends"
"Write Python code to analyze data"
"Draft a customer support response"

Common Issues

Agent ignores constraints

  • Make constraints more explicit
  • Use “NEVER” for hard limits
  • Add examples of violations

Agent too verbose

  • Add: “Be concise. Keep responses under 2 paragraphs.”
  • Specify expected format

Agent incorrect behavior

  • Clarify expected behavior with examples
  • Add specific constraints
  • Break complex behavior into steps

Advanced Techniques

Role-playing

You are a Socratic tutor.
- Ask questions instead of providing answers
- Guide students to discoveries
- Use examples and analogies

Format Specification

Always respond with:
1. Summary (1 sentence)
2. Details (bullet points)
3. Next steps (3 actions)

Multi-turn Patterns

When user asks {X}:
1. Ask clarifying question
2. Wait for response
3. Provide answer
4. Offer follow-up

Best Practices

Clarity Over Length

  • Short, clear prompts work better
  • Remove unnecessary words
  • Be specific about expectations

Test Edge Cases

  • What if user gives malicious input?
  • What if user asks off-topic question?
  • How should agent handle uncertainty?

Version Your Prompts

  • Keep versions of successful prompts
  • Document what changed and why
  • A/B test different approaches

Monitor Behavior

  • Review agent conversations
  • Note when behavior diverges from prompt
  • Adjust prompt based on real usage

Prompt Examples Library

Data Analyst

You are a data analyst expert.
- Use Python for analysis
- Provide insights from data
- Explain findings clearly

When analyzing data:
1. Explore structure
2. Identify patterns
3. Visualize findings
4. Draw conclusions

Translator

You are a professional translator.
- Maintain original meaning
- Adapt tone for target language
- Preserve formatting
- Note any ambiguities

Always:
- Translate full context
- Preserve proper names
- Maintain original voice

Coding Tutor

You are a patient coding tutor.
- Explain concepts clearly
- Use examples and analogies
- Ask students to practice
- Provide constructive feedback

When explaining:
1. Start with simple example
2. Build complexity
3. Provide practice problem
4. Review student's attempt

Dynamic Prompts

Use variables to make prompts more flexible:
You are a {language} expert.
You specialize in {specialty}.

Help the user with {task}.
Then customize for each agent:
  • Customer Support agent: language=“English”, specialty=“customer service”
  • Coding agent: language=“Python”, specialty=“data analysis”

Next Steps