prompt-engineering

Prompt Engineering

Start here: what are you doing?

Task	Go to
Writing a prompt from scratch	Crafting prompts
Reviewing or debugging an existing prompt	Prompt review checklist
Building an agentic system or multi-step workflow	Agentic prompt patterns
Need deeper detail on a specific technique	Going deeper — pick a source doc and WebFetch it

Crafting prompts

The one principle that matters most: explain why, not just what. Claude reasons better with context than rigid rules. Instead of NEVER use ellipses, write Your response will be read aloud by a text-to-speech engine, so never use ellipses since the TTS engine can't pronounce them. Claude generalizes from the reasoning.

Structure for any non-trivial prompt:

1. Role/context — who is Claude, what's the situation
2. Input — wrapped in descriptive XML tags (<code>, <document>, <query>)
3. Task — what to do, with success criteria
4. Output format — show one concrete example rather than describing the format in prose
5. Use {{variables}} for dynamic content to make prompts reusable and testable

When adding examples (few-shot): use 3–5 diverse examples wrapped in <example> tags. Cover edge cases. Ask yourself: would a pattern-matcher pick up unintended shortcuts from these examples?

Prompt review checklist

Run through this when reviewing or debugging a prompt:

• Over-constraining? — Too many MUST/NEVER/ALWAYS makes prompts brittle. If you see a wall of rigid rules, rewrite them as reasoned guidance. One explained rationale outperforms five shouted commands.
• Kitchen-sink? — If the prompt handles discovery, analysis, and output in one block, break it into a chain. Each stage should have a clear input and output.
• Missing examples? — One concrete input/output example beats a paragraph of format description. If the prompt describes a format without showing it, add an example.
• Telling what NOT to do? — Reframe as positive instructions. "Don't use markdown" → "Write in flowing prose paragraphs." Claude steers better toward a destination than away from one.
• Leftover aggressive language from older models? — Phrases like "CRITICAL: You MUST use this tool" or "If in doubt, use [tool]" cause overtriggering on current Claude models. Dial back to normal phrasing: "Use this tool when..."
• Prefills on the last assistant turn? — Deprecated in Claude 4.6. Use direct instructions, structured outputs, or XML output tags instead. See migration patterns in the prompting best-practices doc listed under Going deeper.
• Vague action language? — "Can you suggest changes?" makes Claude suggest rather than act. "Change this function to..." makes it act. Be explicit about whether you want analysis or action.