AgentSkillsCN

prompt-engineer

优化LLM的提示。适用于编写系统提示或提升代理性能时使用。

SKILL.md
--- frontmatter
name: prompt-engineer
description: Optimize prompts for LLMs. Use when crafting system prompts or improving agent performance.

Prompt Engineering

Craft effective prompts for LLM applications.

When to Use

  • Creating system prompts
  • Improving AI output quality
  • Building AI agents
  • Optimizing token usage
  • Designing prompt templates

Core Techniques

Role Setting

code
You are an expert [role] with [X] years of experience in [domain].
Your task is to [specific goal].

Chain of Thought

code
Think through this step by step:
1. First, analyze [aspect 1]
2. Then, consider [aspect 2]
3. Finally, determine [conclusion]

Show your reasoning before giving the final answer.

Few-Shot Examples

code
Here are examples of the expected format:

Input: [example 1 input]
Output: [example 1 output]

Input: [example 2 input]
Output: [example 2 output]

Now process this input:
Input: {user_input}
Output:

Structured Output

code
Respond in the following JSON format:
{
  "analysis": "your analysis here",
  "confidence": 0.0-1.0,
  "recommendations": ["item1", "item2"]
}

Return valid JSON only, no additional text.

Prompt Templates

Code Review

code
You are a senior code reviewer. Review the code for:
1. Security vulnerabilities
2. Performance issues
3. Code quality and readability
4. Best practices violations

For each issue:
- Severity: Critical/High/Medium/Low
- Location: file:line
- Issue: description
- Fix: suggested solution

Code to review:
{code}

Data Extraction

code
Extract the following information from the text:
- Name: person's full name
- Email: email address
- Company: organization name
- Role: job title

If information is not found, use "NOT_FOUND".
Return as JSON.

Text:
{text}

Classification

code
Classify the following text into one of these categories:
- POSITIVE
- NEGATIVE
- NEUTRAL

Consider tone, sentiment, and overall message.
Respond with only the category name.

Text: {text}
Category:

Best Practices

PracticeDoDon't
InstructionsBe specific and explicitBe vague
FormatSpecify output formatAssume format
ExamplesInclude 2-3 examplesZero-shot for complex
ConstraintsSet clear boundariesLeave open-ended
LengthSet max length if neededAllow unlimited

Testing Prompts

  1. Test with edge cases
  2. Try adversarial inputs
  3. Check consistency across runs
  4. Measure output quality
  5. Track token usage

Examples

Input: "Create a prompt for summarization" Action: Design prompt with length constraint, key points extraction, format spec

Input: "Improve this prompt's output" Action: Add examples, clarify instructions, specify format, test iterations