Prompt Engineering for Claude Agents

The Three Laws of Agent Prompts

1. Right Altitude (Goldilocks Zone)

Test: Can you imagine 3 different valid interpretations? Too high. Can you imagine it breaking on a valid input? Too low.

2. Few-Shot Examples (3-5 Canonical Cases)

Don't list every possibility. Show representative examples:

## Examples

### Example 1: Clear Success Case
Input: [Typical, well-formed input]
Output: [Expected response with reasoning shown]

### Example 2: Edge Case
Input: [Unusual but valid input]
Output: [How to handle gracefully]

### Example 3: What NOT to Do
Input: [Tricky input that could mislead]
Output: [Why the naive approach is wrong, correct approach]

3. Explicit Constraints

Tell Claude what NOT to do:

## Constraints

- Do NOT modify files outside the specified directory
- Do NOT commit changes without confirmation
- NEVER include secrets in outputs
- ALWAYS validate input before processing

Prompt Structure Template

# [Agent/Skill Name]

[One-sentence purpose]

## Context

[Background information Claude needs]

## Responsibilities

1. **[Verb] [Object]**: [Brief description]
2. **[Verb] [Object]**: [Brief description]

## Workflow

When given [input type]:
1. [First action]
2. [Second action]
3. [Third action]

## Examples

### Example 1: [Scenario Name]
**Input**: [Sample]
**Output**: [Expected result]
**Reasoning**: [Why this is correct]

### Example 2: [Edge Case]
**Input**: [Tricky sample]
**Output**: [Correct handling]
**Reasoning**: [Common mistake avoided]

## Anti-Patterns

- **Don't**: [Bad practice]
  **Instead**: [Good practice]

## Output Format

[Specify exact format if needed: JSON, markdown, etc.]

Context Engineering

Tell Claude About Session State

## Session Context

- Context may be compacted between turns
- Important decisions will be saved to CLAUDE.md
- You may need to resume work from a previous session
- Use episodic memory to recall past decisions

Memory Management Hints

## Memory Usage

When working on large tasks:
1. Save progress incrementally to files
2. Document key decisions in CLAUDE.md
3. Use clear commit messages for future reference
4. Create TODO items for incomplete work

Model-Specific Considerations

Haiku (Fast, Cheap)

• Keep prompts concise
• Use for simple, well-defined tasks
• Avoid requiring complex reasoning chains

Sonnet (Balanced)

• Standard detail level
• Good for most agent tasks
• Can handle moderate complexity

Opus (Maximum Capability)

• Worth extra context for complex tasks
• Best for nuanced decisions
• Use for critical, high-stakes work

Prompt Optimization Techniques

1. Front-Load Critical Information

## CRITICAL: [Most important instruction]

[Less critical context below]

2. Use Structured Sections

## Input Format
[What Claude receives]

## Output Format
[What Claude must produce]

## Processing Rules
[How to transform input to output]

3. Provide Escape Hatches

If you encounter a situation not covered by these instructions:
1. State what you're uncertain about
2. Explain your best judgment
3. Ask for clarification if needed

4. Include Verification Steps

Before completing:
- [ ] Verify output matches expected format
- [ ] Check for security concerns
- [ ] Ensure changes are minimal and focused

Common Mistakes

Mistake 1: Overly Prescriptive

# BAD
If the error contains "undefined" and the file is JavaScript:
  Check for null pointer
Else if the error contains "type" and the file is TypeScript:
  Check for type mismatch
...

# GOOD
Diagnose errors by:
1. Reading the error message carefully
2. Identifying the error category (type, null, async, etc.)
3. Locating the source in the codebase
4. Understanding the root cause before fixing

Mistake 2: No Examples

# BAD
Review code for quality issues.

# GOOD
Review code for quality issues.

Example issue types:
- Unused variables (remove them)
- Complex conditions (extract to named functions)
- Missing error handling (add try/catch or validation)

Mistake 3: Conflicting Instructions

# BAD
Be thorough and check everything.
Also be fast and efficient.

# GOOD
Prioritize checks:
1. Critical: Security vulnerabilities (always check)
2. Important: Logic errors (check if time allows)
3. Nice-to-have: Style issues (skip if time-constrained)

Testing Your Prompts

Adversarial Testing

Try inputs designed to break the prompt:

• Empty input
• Malformed input
• Edge cases
• Unexpected formats

A/B Testing

Compare prompt variations:

1. Run both on same inputs
2. Evaluate which produces better results
3. Iterate on winner

User Feedback Loop

1. Deploy prompt
2. Collect failure cases
3. Add examples for failures
4. Repeat