Prompt Engineering

Craft effective prompts for LLMs using proven techniques.

When to Use

• Writing system prompts for agents
• Designing few-shot examples
• Structuring outputs from LLMs
• Improving prompt quality

Core Techniques

Role Prompting

You are a senior security engineer reviewing code for vulnerabilities.
Focus on OWASP Top 10 issues. Be specific about line numbers.

Chain-of-Thought

Think through this step by step:
1. What is the input?
2. What rules apply?
3. What is the output?

Few-Shot

Convert these natural language queries to SQL:

Q: "Show all users who signed up this week"
A: SELECT * FROM users WHERE created_at >= DATE('now', '-7 days')

Q: "Find orders over $100"
A: SELECT * FROM orders WHERE amount > 100

Q: "List active subscriptions"
A: 

Structured Output

Respond in this exact JSON format:
{
  "summary": "one line summary",
  "issues": [{"severity": "high|medium|low", "file": "path", "description": "..."}],
  "recommendation": "what to do"
}

Anthropic Best Practices

1. Be specific — "Write a Python function that..." not "Write some code"
2. Use XML tags — <instructions>, <example>, <context> for structure
3. Provide examples — Show, don't just tell
4. Set constraints — "Respond in under 100 words"
5. Use delimiters — Separate context from instructions clearly

Anti-Patterns

Output Control

# Force specific format
Respond ONLY with valid JSON. No explanation, no markdown.

# Limit length
In exactly 3 bullet points...

# Require citations
Cite sources with [Source: URL] format.