Claude Prompt Patterns

A cheat sheet of patterns for writing effective Claude prompts, based on Anthropic guidance as of spring 2026.

• XML tags as semantic boundaries between prompt sections (context, task, rules, schema)
• Include the JSON schema directly rather than describing the shape you want
• Positive framing over "do not" — tell the model what to do, not what to avoid
• Explain why constraints exist — Claude generalizes from motivation better than from bare rules
• Few-shot examples are the highest-leverage fix for inconsistent output
• Data first, task after in long prompts for better attention
• System prompt for persona and persistent constraints; user prompt for per-call context and data