prompt-engineering

Prompt Engineering for Agents

Guidelines for writing effective system prompts and agent instructions.

When to Apply

• Reviewing or improving agent system prompts
• Writing instructions for new agents or tools
• Optimizing prompts based on evaluation results
• Analyzing prompt effectiveness

Quick Reference: The Four Principles

Principle	Do	Don't
Right Altitude	Principle-based guidance	Rigid Step 1,2,3 workflows
RFC 2119 Keywords	MUST/SHOULD/MAY for precision	Vague "you should" everywhere
Structure Over Prescription	When-then patterns	Complex if-else chains
Examples Over Rules	Concrete demonstrations	Abstract prescriptions

RFC 2119 Keywords

Keyword	Meaning	Use For
MUST	Always required	Security, correctness
MUST NOT	Absolute prohibition	Dangerous operations
SHOULD	Strong guidance, exceptions OK	Best practices
MAY	Optional, agent discretion	Optimizations

System Prompt Template

# Role

[What the agent is and primary function]

# Constraints

[Hard boundaries using RFC 2119]

# Tools & Integration

[When and how to use available tools]

# Decision Framework

[Principles for making choices]

# Examples

[Concrete demonstrations]

How It Works

1. Read - Access current prompt via read_file
2. Analyze - Check for over-prescription, under-specification, conflicts
3. Improve - Apply principles: workflows → principles, add RFC 2119, add examples
4. Document - Record what changed and why
5. Test - Run against evaluation, compare results

Common Anti-Patterns

Anti-Pattern	Example	Fix
Prompt bloat	Every scenario in prompt	Extract to skills, keep principles
Conflicting guidance	"Be creative and follow exact steps"	Prioritize with RFC 2119
Assuming context	"Use best practices"	Specify what that means
Workflow addiction	"1. Read, 2. Edit, 3. Test"	"Understand context before modifying"

Best Practices

• Balance: not too rigid (can't adapt), not too vague (no guidance)
• Use RFC 2119 to distinguish hard requirements from suggestions
• Provide examples showing desired behavior, not just rules
• Separate concerns: role, constraints, tools, decisions, examples
• Test changes with evaluation harness before committing