how-to-write-prompts-claude-engineer

name	how-to-write-prompts-claude-engineer
description	Anthropic engineers' best practices for prompt writing. Use when improving prompts, teaching prompt engineering, or understanding model interaction. Trigger phrases: "Claude工程师", "提示词技巧", "不要角色扮演", "实话实说"

How to Write Prompts - Claude Engineer Insights

Overview

Anthropic prompt engineers' key principles: clear communication, iteration, honesty with models, no unnecessary role-play.

Core Principles

1. Clear Communication

• Write very clear task descriptions, not abstract constructs
• Articulate clearly so model understands the task
• Organize everything you know but model doesn't, then write it down

2. Iteration

• Prompt engineering = trial and error process
• Willingness to iterate and observe is critical
• Test edge cases: empty input, unusual data, boundary conditions
• When model errs: ask "Why did you get this wrong? Can you help fix my prompt?"

3. No Role-Play Needed

• As models improve, no need to lie to them
• Don't pretend "I'm a teacher giving a quiz" - say "I'm building an LLM evaluation dataset"
• Imagine a capable temp worker with little context - describe your task directly
• Record yourself explaining the task, transcribe, paste into prompt - often works better

4. Trust Model Capability

• Don't treat model like a child; don't over-simplify
• Models can understand complex information
• Give papers directly instead of simplified "baby version"
• Respect model's ability to handle context

5. When to Give Up

• Some tasks: every adjustment pushes result further from goal → abandon
• Better to wait for next model than spend months on unsolvable prompt

Research vs Consumer Prompts

• Research: Fewer examples, seek diversity, explore model boundaries
• Consumer: Many examples, prioritize stability and consistency
• Enterprise: Must handle millions of uses, consider all edge cases

Usage Example

User input: "我的提示词效果不好，怎么改进？" AI action: Applies principles - clarify task, test edge cases, try transcribing verbal explanation, avoid role-play shortcuts Expected result: Improved prompt with clear task description and edge case handling