Prompt Engineering

Techniques for getting reliable, high-quality output from LLMs.

Core Techniques

System Prompt Design

• Define the role, constraints, and output format upfront
• Be explicit about what NOT to do
• Put the most important instructions first

Chain-of-Thought

Force step-by-step reasoning before the answer:

Think through this step by step:
1. First, identify...
2. Then, evaluate...
3. Finally, decide...

Few-Shot Examples

Provide 2-3 examples of input → output pairs before the actual task. Match the exact format you want back.

Structured Output

Request specific formats:

Respond with JSON in this exact format:
{"decision": "approve|reject", "reason": "string", "confidence": 0.0-1.0}

Prompt Patterns

Persona Pattern

You are a senior security engineer reviewing code for vulnerabilities.
Focus only on security issues. Ignore style, naming, and formatting.

Constraint Pattern

Rules:
- Maximum 3 sentences per response
- Use only information from the provided context
- If uncertain, say "I don't know" instead of guessing

Verification Pattern

After generating your response, verify:
1. Does it answer the original question?
2. Are all claims supported by evidence?
3. Are there any logical contradictions?
If any check fails, revise before responding.

Decomposition Pattern

Break this problem into sub-problems:
1. Solve each sub-problem independently
2. Combine the solutions
3. Verify the combined solution

API Usage Patterns

Claude API

curl -s https://api.anthropic.com/v1/messages \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "claude-sonnet-4-20250514",
    "max_tokens": 1024,
    "system": "You are a code reviewer. Only flag security issues.",
    "messages": [{"role": "user", "content": "Review this code: ..."}]
  }' | jq '.content[0].text'

OpenAI API

curl -s https://api.openai.com/v1/chat/completions \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-4o",
    "messages": [
      {"role": "system", "content": "You are a code reviewer."},
      {"role": "user", "content": "Review this code: ..."}
    ],
    "response_format": {"type": "json_object"}
  }' | jq '.choices[0].message.content'

Prompt Optimization

Reduce token usage

• Remove filler words and redundant instructions
• Use examples instead of lengthy descriptions
• Put context in system prompt (cached across turns with some providers)

Improve reliability

• Ask the model to output reasoning before conclusions
• Use XML tags to delimit sections: <context>...</context>
• Be specific: "list 5 items" not "list some items"
• Provide negative examples: "Do NOT include..."

Test prompts

• Run the same prompt 5 times — if outputs vary wildly, the prompt is ambiguous
• Use temperature 0 for deterministic tasks, 0.7+ for creative tasks
• Compare model outputs across providers for the same prompt

Notes

• Prompt engineering is empirical — test and iterate, don't theorize.
• Long prompts aren't bad if they reduce ambiguity. Clarity > brevity.
• System prompts are more reliable than user-message instructions for persistent behavior.
• XML tags (<instructions>, <context>, <examples>) help models parse structured prompts.
• Cache-friendly prompts: put static content first, dynamic content last.