Best Practices Guide for Prompt Engineering (Beginner-Friendly)
1. Role Explanation
Three roles are involved in prompt application:
- USER: The user who poses questions or gives instructions
- ASSISTANT: The model responsible for generating responses
- SYSTEM: The system role that sets model behavior rules (e.g., defining background logic in API development, or simulating via direct user instructions like “When helping me write, include at least one joke or witty remark in each paragraph”)
2. Best Practice 1: Write Clear Instructions
Core Principle
GPT cannot read users’ minds. Clearly specify details, formats, roles, etc., to reduce model guesswork.
Specific Strategies
1. Include Detailed Information
- Target Audience: Beginners
- Difficulty: ⭐
- Example:Vague Query: “Recommend a book”Clear Query: “Recommend a communication skills book for career newcomers, published in the past 3 years, with a 豆瓣 score above 8”
2. Assign a Model Role
- Target Audience: Beginners
- Difficulty: ⭐
- Example:
3. Use Delimiters to Separate Content
- Target Audience: Beginners
- Difficulty: ⭐
- Example:
4. Break Tasks into Steps
- Target Audience: Beginners
- Difficulty: ⭐
- Example:
5. Provide Example References (Few-Shot Prompting)
- Target Audience: Beginners
- Difficulty: ⭐
- Example:
6. Specify Output Length
- Target Audience: Beginners
- Difficulty: ⭐
- Note: When specifying word counts, use a range (e.g., “50-70 words”). Models better control paragraph/bullet counts.
- Example:
3. Best Practice 2: Provide Reference Texts
Core Principle
Avoid model fabrications by providing credible texts to ensure answer accuracy.
Specific Strategies
1. Answer Based on Reference Texts
- Target Audience: Beginners
- Difficulty: ⭐
- Example:
2. Cite Text Passages
- Target Audience: Beginners
- Difficulty: ⭐
- Example:
4. Best Practice 3: Decompose Complex Tasks
Core Principle
Break complex tasks into simple subtasks to reduce error rates and improve efficiency.
Specific Strategies
1. Intent Classification for Guided Processing
- Target Audience: Beginners/Professionals
- Difficulty: ⭐⭐
- Example:
2. Summarize and Filter Long Conversations
- Target Audience: Professionals/Developers
- Difficulty: ⭐⭐
- Scenario: Regularly summarize medical histories in telemedicine, retaining only information relevant to current symptoms.
3. Segment and Summarize Long Documents
- Target Audience: Professionals
- Difficulty: ⭐⭐
- Method: Summarize paragraph by paragraph → Merge paragraph summaries → Recursively generate a full summary to avoid context limits.
5. Best Practice 4: Allow Model “Thinking” Time
Core Principle
Guide the model to reason step-by-step for complex problems to reduce intuitive errors.
Specific Strategies
1. Reason First, Then Judge
- Target Audience: Professionals/Developers
- Difficulty: ⭐⭐
- Example:
2. Hide Reasoning Process (Inner Monologue)
- Target Audience: Professionals/Developers
- Difficulty: ⭐⭐⭐
- Scenario: In education, the model generates hints through internal reasoning to avoid exposing answers directly.
3. Check for Omissions
- Target Audience: Professionals/Developers
- Difficulty: ⭐⭐⭐
- Example:
6. Best Practice 5: Use External Tools
Core Principle
Leverage external tools to compensate for model limitations (e.g., retrieval, calculation, API calls).
Specific Strategies
1. Embedding-Based Search for Knowledge Retrieval
- Target Audience: Developers
- Difficulty: ⭐⭐⭐⭐
- Principle: Convert text to vectors and match relevant content via cosine similarity (e.g., movie recommendation systems).
2. Code Execution for Complex Calculations
- Target Audience: Professionals/Developers
- Difficulty: ⭐⭐⭐
- Example:
7. Best Practice 6: Systematically Test Changes
Core Principle
Evaluate model outputs through standardized tests to ensure optimization effectiveness.
Specific Strategies
1. Evaluate Against Reference Answers
- Target Audience: Professionals/Developers
- Difficulty: ⭐⭐⭐⭐
- Example:
Conclusion
Mastering prompt engineering best practices can significantly enhance AI interaction efficiency. Key strategies include: clarifying instructions, providing context, decomposing tasks, guiding deep thinking, integrating external tools, and systematic testing. For advanced scenarios, further learning of structured prompting techniques is recommended.