Prompt Engineering

Source Cursor rule: .cursor/rules/prompt-engineering.mdc. Original file scope: .cursor/rules/*.mdc. Original Cursor alwaysApply: false.

Prompt Engineering Best Practices

Based on Claude's Prompt Engineering Documentation

Core Principles

1. Define Success Criteria First

Before writing prompts:

• Establish clear objectives: What constitutes a successful response?
• Create evaluation metrics: How will you measure prompt effectiveness?
• Draft and iterate: Start with a first draft and refine based on results

2. When to Use Prompt Engineering vs Fine-tuning

Prompt engineering is preferred because:

• Resource efficient: Only requires text input, no GPUs
• Cost effective: Uses base model pricing
• Maintains updates: Works across model versions
• Time saving: Instant results vs hours/days for fine-tuning
• Minimal data needs: Works with zero-shot or few-shot
• Flexible iteration: Quick experimentation cycle
• Preserves knowledge: No catastrophic forgetting
• Transparent: Human-readable, easy to debug

The 6 Core Techniques

1. Be Clear and Direct

Principle: Provide explicit, unambiguous instructions.

❌ Bad: "Tell me about it"
✅ Good: "Summarize the following article in three bullet points, focusing on key findings"

❌ Bad: "Help with code"
✅ Good: "Debug this Python function that should return the sum of even numbers in a list"

Tips:

• State the task explicitly at the start
• Specify the desired output format (bullet points, JSON, paragraphs)
• Include constraints (word count, tone, audience)
• Mention what to include AND what to exclude

2. Use Examples (Multishot Prompting)

Principle: Show the model what you want through examples.

<examples>
  <example>
    <input>The movie was absolutely terrible, waste of time</input>
    <output>{"sentiment": "negative", "confidence": 0.95}</output>
  </example>
  <example>
    <input>Decent film, not great but watchable</input>
    <output>{"sentiment": "neutral", "confidence": 0.7}</output>
  </example>
  <example>
    <input>Best movie I've seen this year!</input>
    <output>{"sentiment": "positive", "confidence": 0.9}</output>
  </example>
</examples>

Now analyze: "The special effects were amazing but the plot was confusing"

Tips:

• Include 3-5 diverse examples covering edge cases
• Show examples of BOTH good and bad outputs
• Match example complexity to your actual use case
• Order examples from simple to complex

3. Let Claude Think (Chain of Thought)

Principle: Encourage step-by-step reasoning for complex tasks.

<instruction>
Solve this problem step by step. Show your reasoning before giving the final answer.
</instruction>

<problem>
A train leaves Station A at 9:00 AM traveling at 60 mph. Another train leaves
Station B at 10:00 AM traveling at 80 mph toward Station A. The stations are
280 miles apart. When will the trains meet?
</problem>

<thinking>
[Let Claude work through the problem here]
</thinking>

<answer>
[Final answer after reasoning]
</answer>

Tips:

• Use phrases like "Think step by step" or "Explain your reasoning"
• For complex tasks, explicitly request a thinking section
• Chain of thought improves accuracy on math, logic, and multi-step problems
• Can use <thinking> tags to separate reasoning from output

4. Use XML Tags

Principle: Structure prompts with clear delimiters for better parsing.

<context>
You are helping debug a penetration testing tool that automates security scans.
</context>

<task>
Analyze the following error log and identify the root cause.
</task>

<error_log>
[2024-01-15 10:23:45] ERROR: Connection timeout after 30s
[2024-01-15 10:23:45] DEBUG: Target: 192.168.1.1:443
[2024-01-15 10:23:45] DEBUG: Retry attempt 3 of 3
</error_log>

<output_format>
Provide your analysis in this format:
- Root cause: [one sentence]
- Evidence: [relevant log lines]
- Recommended fix: [actionable steps]
</output_format>

Common XML Tags:

• <context> - Background information
• <task> or <instruction> - What to do
• <examples> - Sample inputs/outputs
• <constraints> - Limitations or rules
• <output_format> - Expected response structure
• <thinking> - Reasoning section
• <answer> - Final response

5. Give Claude a Role (System Prompts)

Principle: Assign a persona to influence response style and expertise.

<role>
You are a senior security researcher with 15 years of experience in penetration
testing. You specialize in web application security and have discovered multiple
CVEs. You communicate findings clearly and prioritize actionable recommendations.
</role>

<task>
Review this HTTP response and identify potential security vulnerabilities.
</task>

Effective Role Elements:

• Expertise level (senior, expert, specialist)
• Domain knowledge (security, finance, medicine)
• Communication style (technical, friendly, formal)
• Priorities (accuracy, brevity, thoroughness)

6. Prefill Claude's Response

Principle: Start the response to guide format and direction.

Human: List the top 3 security vulnerabilities in this code.