optimizing-llm-prompts

kylehughes 的头像作者 kylehughes

完善和结构化LLM的提示,以确保清晰度、可靠性和最佳性能。在编写系统提示、复杂指令或调试代理行为时使用。

📖 中文导读:完善和结构化LLM的提示,以确保清晰度、可靠性和最佳性能。在编写系统提示、复杂指令或调试代理行为时使用。 以下正文为该技能的原始 SKILL.md 内容(英文),上方安装命令可直接使用。

name: optimizing-llm-prompts description: Refine and structure prompts for LLMs to ensure clarity, reliability, and optimal performance. Use when writing system prompts, complex instructions, or debugging agent behaviors.

Optimizing LLM Prompts

Instructions

Follow these steps to create robust and effective prompts for LLMs (specifically Claude).

  1. Define the Goal: Clearly identify the desired output and behavior. Be specific about format, tone, and constraints.
  2. Structure with XML: Use XML tags to delineate sections.
    • <system>: High-level role and identity.
    • <context>: Static background information.
    • <rules>: Specific constraints and instructions.
    • <examples>: Few-shot demonstrations.
  3. Draft Instructions:
    • Use imperative voice ("Do this", not "You should").
    • Quantify everything (e.g., "3 sentences" not "concise").
    • Use positive framing (what to do, not just what not to do).
  4. Add Examples: Provide 1-3 examples of input -> output mapping to "show" the model what you want.
  5. Iterate: Test with edge cases. If the model fails, add a specific rule or example to address that failure mode.

Best Practices Summary

  • XML Structure: Essential for Claude to distinguish between instructions and data.
  • Chain of Thought: Ask the model to "think step-by-step" before answering complex queries.
  • Progressive Disclosure: Don't dump all context; allow the model to request more if needed.
  • Input Sanitation: Wrap user input in distinct tags (e.g., <user_query>) to prevent prompt injection.

Checklist

  • Structure: Are sections clearly separated (XML/Headers)?
  • Clarity: Are instructions imperative and quantified?
  • Safety: Is there a catch-all override for conflicting user requests?
  • Examples: Are there few-shot examples for complex behaviors?
  • Context: Is static context separated from dynamic user queries?
  • Output: Is the output format explicitly defined (JSON, Markdown, etc.)?

Detailed Guidance

For a deep dive on critical rules, forbidden practices, and optimization patterns, see REFERENCE.md.