prompt-engineering

Apply these techniques when creating prompts, agents, commands, or system instructions.

Core Principles

• Treat Claude as context-free: Provide complete information
• Be explicit: Don't say "Analyze this" - say "Analyze for X, Y, Z risks with ratings"
• Explain WHY: Tell motivation, not just what to do
• Define success criteria: Specify what "good" looks like

XML Tags for Structure

Use XML tags to separate prompt components:

<context>Background and rationale</context>
<instructions>
1. Specific task requirements
2. Edge case handling
</instructions>
<examples>
  <example>
    <input>Sample input</input>
    <output>Expected output</output>
  </example>
</examples>
<data>{{VARIABLE_DATA}}</data>

Multishot Prompting (Few-Shot)

Provide 3-5 examples for complex tasks:

• Examples teach format AND correct behavior
• Show edge cases and variations
• Quality over quantity: one excellent > three mediocre
• Use consistent formatting across examples

Best for: JSON/XML generation, classifications, style matching

Chain-of-Thought

For complex reasoning, request step-by-step thinking:

<thinking>
Step-by-step reasoning process
</thinking>
<answer>
Final conclusion
</answer>

Best for: Arithmetic, logic, multi-step analysis, decisions

Response Prefilling

Guide output format by starting the assistant's reply:

messages=[
  {"role": "user", "content": "Analyze this data: {{DATA}}"},
  {"role": "assistant", "content": "{\n  \"analysis\":"}
]

Claude 4-Specific Patterns

| Behavior | How to Request | |----------|----------------| | Comprehensive output | "Include as many relevant features as possible" | | Action vs suggestion | "Change this..." vs "Can you suggest..." | | Summaries | "After completing, provide a quick summary" |

System Prompts

Use system parameter for role/behavior, user for tasks:

system="""You are a senior solutions architect.

Communication Style:
- Be concise and technical
- Provide concrete examples

Constraints:
- Never speculate without data
- Recommend industry-standard solutions first"""

Anti-Patterns to Avoid

• Assuming shared knowledge
• Using vague descriptors ("be creative")
• Leaving format unspecified
• Telling what NOT to do (use positive instructions)
• Relying on implications

Integration Pattern

1. Start with clear, direct instructions
2. Add structure with XML tags for complex prompts
3. Provide examples via multishot for format/style
4. Elicit reasoning with CoT for complex problems
5. Use prefilling to enforce specific output formats

Long Context Tips

• Place long documents at TOP of context
• Put queries and instructions at BOTTOM
• Use prompt caching for frequently reused context