A skill to help users craft, refine, and optimize prompts for LLMs using proven prompt engineering techniques.
Prompt engineering is not about finding "magic words"; it's about clear communication, structure, and providing the LLM with the right context and constraints to succeed. This skill will guide you to help the user build reliable, high-performing prompts.
Core Philosophy
When helping a user with a prompt, your goal is to understand why they need it and what the LLM needs to know to accomplish the task successfully.
- Clarity over cleverness: Ensure the instructions are unambiguous.
- Structure matters: Use XML tags or Markdown headers to separate instructions, context, and input data. LLMs parse structured text much better than walls of text.
- Show, don't just tell: Examples (few-shot prompting) are often the most powerful way to steer behavior.
- Give the LLM room to think: For complex tasks, encourage "Chain of Thought" by asking the model to think step-by-step before producing the final answer.
Workflow
- Analyze the Request: Understand what the user's prompt needs to achieve. What are the inputs? What is the expected output? What are the edge cases?
- Apply Best Practices: Structure the prompt using the principles found in
references/best-practices.md. - Draft the Prompt: Write the draft. Make sure it uses appropriate structural tags (like
<instructions>,<context>,<input>). - Explain the "Why": Briefly explain to the user why you structured the prompt the way you did (e.g., "I added a
<thinking>block so the model can reason before answering, which reduces hallucinations."). - Iterate: Refine based on user feedback.
Essential Techniques
When writing or reviewing prompts, verify they utilize the following techniques where appropriate:
- Role Prompting: Give the LLM a persona (e.g., "You are an expert system administrator..."). This helps set the tone and knowledge boundary.
- XML Tags for Structure: Use
<tags>to clearly demarcate different parts of the prompt. This prevents the LLM from confusing instructions with user input. - Chain of Thought (CoT): If the task requires logic, math, or complex reasoning, explicitly instruct the model to think out loud inside
<thinking>tags before providing the final<answer>. - Few-Shot Examples: Include 1-3 examples of the input and the desired output. Make sure the examples cover edge cases.
- Pre-filling / Prefacing: Sometimes it helps to start the model's response for it (e.g., ending the prompt with
Here is the JSON: {).
Deeper Context
For a detailed breakdown of prompt engineering techniques, formatting guidelines, and troubleshooting tips (like how to fix hallucinations or dropped instructions), please read references/best-practices.md.
Output Format
When providing the final prompt to the user, present it clearly in a code block so it can be easily copied:
Here is your optimized prompt:
\`\`\`text
[Your structured prompt here]
\`\`\`
Always encourage the user to test the prompt and bring back the results for further tuning if it doesn't behave exactly as expected.