💬 Prompting Patterns

🎓 Page Overview

Trang này cung cấp các patterns và best practices cho prompt engineering, từ system prompts đến few-shot learning và cách tránh các anti-patterns phổ biến.

Level: Core Solves: Thiết kế prompts hiệu quả cho production LLM applications với consistency và reliability

🎯 System Prompt Engineering

Anatomy of a System Prompt

# Role Definition
You are [ROLE] with expertise in [DOMAIN].
Your purpose is to [PRIMARY_OBJECTIVE].

# Behavioral Guidelines
- Always [POSITIVE_BEHAVIOR]
- Never [PROHIBITED_BEHAVIOR]
- When uncertain, [UNCERTAINTY_HANDLING]

# Output Format
Respond in [FORMAT] with the following structure:
[STRUCTURE_DEFINITION]

# Constraints
- Maximum response length: [LIMIT]
- Language: [LANGUAGE]
- Tone: [TONE]

# Examples (Optional)
[FEW_SHOT_EXAMPLES]

System Prompt Components

Component	Purpose	Example
Role Definition	Establish persona và expertise	"You are a senior software engineer reviewing code"
Task Framing	Define primary objective	"Help users debug and optimize their code"
Behavioral Rules	Set guardrails	"Never provide code that could cause security vulnerabilities"
Output Format	Standardize responses	"Respond in JSON with keys: analysis, suggestion, code"
Constraints	Limit scope	"Focus only on Python; redirect other languages politely"

📐 Prompt Structure Patterns

1. Instruction-Input-Output (IIO) Pattern

## Instruction
Analyze the following customer feedback and extract:
1. Sentiment (positive/negative/neutral)
2. Key topics mentioned
3. Action items if any

## Input
{customer_feedback}

## Output Format
Return JSON:
{
  "sentiment": "...",
  "topics": [...],
  "action_items": [...]
}

2. Chain-of-Thought (CoT) Pattern

Solve this step by step:

1. First, identify the core problem
2. Then, analyze the constraints
3. Consider possible approaches
4. Select the best approach with reasoning
5. Provide the final solution

Problem: {problem_description}

Show your reasoning at each step.

3. Role-Task-Format (RTF) Pattern

# Role
You are an expert data analyst at a Fortune 500 company.

# Task
Analyze the sales data and provide insights on:
- Trends over the past quarter
- Top performing products
- Recommendations for next quarter

# Format
Structure your response as:
1. Executive Summary (2-3 sentences)
2. Key Findings (bullet points)
3. Recommendations (numbered list)
4. Supporting Data (table format)

🎓 Few-Shot Learning Patterns

Static Few-Shot

Extract entities from the text.

Example 1:
Text: "Apple announced new iPhone 15 in Cupertino on September 12"
Entities: {"company": "Apple", "product": "iPhone 15", "location": "Cupertino", "date": "September 12"}

Example 2:
Text: "Microsoft CEO Satya Nadella spoke at Build 2024 in Seattle"
Entities: {"company": "Microsoft", "person": "Satya Nadella", "event": "Build 2024", "location": "Seattle"}

Now extract from:
Text: "{input_text}"
Entities:

Dynamic Few-Shot (Retrieval-Augmented)

Implementation Pattern:

1. Maintain example bank with embeddings
2. Retrieve semantically similar examples
3. Select diverse, representative examples
4. Include in prompt with recency/relevance weighting

Few-Shot Best Practices

Practice	Rationale
3-5 examples	Enough diversity without overwhelming context
Diverse examples	Cover edge cases và variations
Ordered by complexity	Simple → Complex for gradual learning
Include negative examples	Show what NOT to do
Match target distribution	Examples should reflect real queries

🚫 Prompt Anti-Patterns

1. Ambiguous Instructions

❌ Bad:

Help me with this data.

✅ Good:

Analyze this CSV data and:
1. Calculate the average sales per region
2. Identify the top 3 performing products
3. Return results as a markdown table

2. Overloaded Prompts

❌ Bad:

You are an expert in everything. Help with any question about coding, 
writing, math, science, history, cooking, legal advice, medical diagnosis...

✅ Good:

You are a Python code reviewer specializing in:
- Code quality and PEP8 compliance
- Performance optimization
- Security best practices

Redirect other topics with: "This is outside my expertise."

3. Missing Output Format

❌ Bad:

Extract the information from this document.

✅ Good:

Extract information and return as JSON:
{
  "title": "string",
  "date": "YYYY-MM-DD",
  "key_points": ["string"],
  "confidence": 0.0-1.0
}

4. Inconsistent Examples

❌ Bad:

Example 1: Return "positive" or "negative"
Example 2: Return {"sentiment": "good", "score": 0.8}

✅ Good:

Example 1: {"sentiment": "positive", "confidence": 0.95}
Example 2: {"sentiment": "negative", "confidence": 0.87}

🔧 Advanced Techniques

Prompt Chaining

Self-Consistency

Solve this problem 3 times with different approaches.
Then compare your answers and provide the most likely correct answer.

Problem: {problem}

Approach 1: [reasoning and answer]
Approach 2: [reasoning and answer]
Approach 3: [reasoning and answer]

Final Answer: [most common/confident answer with explanation]

Constitutional AI Pattern

# Base Instruction
{original_instruction}

# Constitutional Principles
Before responding, verify your answer against:
1. Accuracy: Is the information factually correct?
2. Helpfulness: Does this actually help the user?
3. Safety: Could this cause harm?
4. Honesty: Am I being transparent about limitations?

If any principle is violated, revise your response.

📋 Prompt Engineering Checklist

Design Phase

• [ ] Define clear objective and success criteria
• [ ] Identify required output format
• [ ] Prepare diverse few-shot examples
• [ ] Document edge cases và expected behaviors

Testing Phase

• [ ] Test with diverse inputs
• [ ] Verify output format consistency
• [ ] Check handling of edge cases
• [ ] Measure latency impact of prompt length

Production Phase

• [ ] Version control prompts
• [ ] A/B test prompt variations
• [ ] Monitor output quality metrics
• [ ] Set up prompt injection detection

🔗 Cross-References

• 📎 RAG Engineering - Combining prompts with retrieved context
• 📎 LLM Evaluation - Measuring prompt effectiveness
• 📎 LLM Safety - Prompt injection defense

📚 Further Reading

• "The Prompt Engineering Guide" - DAIR.AI
• "Anthropic Prompt Engineering" - Anthropic Documentation
• "OpenAI Best Practices" - OpenAI Cookbook