--- name: prompt-engineer-prompt-patterns description: "Comprehensive reference for prompt engineering patterns and techniques. 提示工程模式與技術全覽。 Use when: selecting prompting strategy, applying zero/few-shot or CoT patterns, structuring prompts for specific task types." disable-model-invocation: true --- # Prompt Engineering Patterns Reference (2026) 汝為提示工程專家。此参考供設計、分析、優化提示之用。 ## Foundational Techniques ### Zero-Shot Prompting 無示例直接指令。適用能力強之模型處理明確任務。 ``` Classify the sentiment of this review as positive, negative, or neutral: Review: "The product exceeded my expectations!" Sentiment: ``` **When to use**：簡單任務、能力強之模型、輸出格式明確 **When to avoid**：複雜推理、模糊任務、格式敏感輸出 ### Few-Shot Prompting 正式任務前提供 3-5 個多樣示例。 ``` Classify the sentiment: Review: "Absolutely loved it!" -> Positive Review: "Total waste of money" -> Negative Review: "It works as expected" -> Neutral Review: "Could be better but not bad" -> ``` **Best practices**： - 使用 3-5 個多樣高品質示例 - 示例須與期望輸出格式完全一致 - 涵蓋邊界案例 - 由簡至繁排序 **Note on reasoning models**：少樣本示例常**降低** DeepSeek R1、OpenAI o1/o3 等模型效能。 ### Chain-of-Thought (CoT) Prompting 鼓勵逐步推理後作答。 ``` Question: If a store has 47 apples and sells 23, then receives a shipment of 35, how many apples does it have? Let me work through this step by step: 1. Starting apples: 47 2. After selling 23: 47 - 23 = 24 3. After receiving 35: 24 + 35 = 59 Answer: 59 apples ``` **Trigger phrases**： - "Let's think step by step" - "Walk me through your reasoning" - "Show your work" **2026 研究發現**：CoT 效果隨新模型遞減。對內建推理模型（o1、o3、R1），CoT 提示僅帶微小增益（2.9-3.1%），卻增加 20-80% 回應時間。 **When to use CoT**：標準 LLM 處理複雜多步問題 **When to avoid**：推理模型、簡單任務、時間敏感應用 ### Self-Consistency 生成多條推理路徑，選最一致答案。 ``` Generate 3 different approaches to solve this problem, then select the answer that appears most frequently. ``` **Best for**：數學問題、邏輯推理、高風險決策 ### Tree of Thoughts (ToT) 系統性探索多推理分支。 ``` Consider this problem from multiple angles: Branch 1: [Approach A] - Step 1.1: ... - Step 1.2: ... - Evaluation: [Pros/cons] Branch 2: [Approach B] - Step 2.1: ... - Step 2.2: ... - Evaluation: [Pros/cons] Best path: [Selected branch with justification] ``` **Best for**：複雜決策、策略問題、創意探索 ## Advanced Techniques ### ReAct (Reasoning + Acting) 推理與行動交替進行。 ``` Question: What is the population of the capital of France? Thought: I need to find the capital of France first. Action: Search "capital of France" Observation: Paris is the capital of France. Thought: Now I need to find Paris's population. Action: Search "population of Paris" Observation: Paris has approximately 2.1 million inhabitants. Thought: I have the answer. Answer: The population of Paris, the capital of France, is approximately 2.1 million. ``` **Best for**：工具使用代理、信息收集、多步任務 ### Reflexion 基於反饋迭代精煉回應。 ``` [First attempt] What could be improved: - [Issue 1] - [Issue 2] [Refined answer addressing issues] ``` ### Directional Stimulus Prompting (DSP) 提供提示而不給出答案。 ``` Solve this math problem. Hint: Consider using the quadratic formula. Problem: x² + 5x + 6 = 0 ``` ### Program-Aided Language Models (PAL) 生成程式碼解決問題。 ``` Solve this problem by writing Python code: Problem: Calculate the compound interest on $10,000 at 5% annual rate for 3 years. ```python principal = 10000 rate = 0.05 time = 3 compound_interest = principal * (1 + rate)**time - principal print(f"Compound interest: ${compound_interest:.2f}") ``` ``` ## Structural Patterns ### Role-Task-Tone Framework 以明確角色、任務、語調結構化提示。 ``` Role: You are a senior software architect with 15 years of experience in distributed systems. Task: Review this microservices architecture design and identify potential bottlenecks, single points of failure, and scalability concerns. Tone: Be direct and technical. Prioritize actionable feedback over praise. ``` ### XML Delineation 以 XML 標籤結構化複雜提示（對 Claude 尤為有效）。 ```xml Background information and constraints Clear statement of what to accomplish Expected output structure Sample input Expected output ``` ### Markdown Structuring 以標題及格式提升可讀性。 ```markdown ## Context [Background information] ## Task [What to do] ## Requirements - Requirement 1 - Requirement 2 ## Output Format [Structure specification] ``` ## Output Control Patterns ### Structured Output Forcing 強制指定輸出格式。 ``` Return your response as valid JSON with this schema: { "summary": "string (max 100 words)", "key_points": ["string"], "confidence": "number (0-1)" } ``` ### Template Fill-in 提供精確模板供填寫。 ``` Complete this template: Summary: [1-2 sentences] Main Argument: [Single sentence] Supporting Evidence: [3 bullet points] Conclusion: [1 sentence] ``` ### Positive Framing 陳述應做之事，而非應避之事。 **Instead of**: "Don't use markdown" **Use**: "Write in flowing prose paragraphs without formatting" **Instead of**: "Never make up information" **Use**: "Only use information from the provided context. If unsure, state your uncertainty." ## Task-Specific Patterns ### Code Generation ``` Write a [language] function that: Purpose: [What it does] Inputs: [Parameters with types] Outputs: [Return value with type] Constraints: [Performance, style requirements] Include: - Error handling for [edge cases] - Comments for complex logic - Type hints/annotations ``` ### Text Summarization ``` Summarize this text in [X words/sentences]: Focus on: [Key aspects to capture] Exclude: [What to omit] Style: [Formal/casual/technical] Format: [Paragraph/bullets/structured] Text: [Content to summarize] ``` ### Analysis/Evaluation ``` Analyze [subject] using these criteria: 1. [Criterion 1]: [What to evaluate] 2. [Criterion 2]: [What to evaluate] 3. [Criterion 3]: [What to evaluate] For each criterion: - Rate on scale of [1-5] - Provide specific evidence - Suggest improvements Output as structured report with executive summary. ``` ## Anti-Patterns to Avoid ### Vague Instructions ❌ "Make it better" ✅ "Improve readability by using shorter sentences and active voice" ### Conflicting Requirements ❌ "Be concise but include all details" ✅ "Prioritize the top 3 most important points" ### Assumed Context ❌ "Like we discussed before..." ✅ "Given that [explicit context]..." ### Negative Framing ❌ "Don't be verbose" ✅ "Keep responses under 200 words" ### Over-Prescription for Reasoning Models ❌ "First do step 1, then step 2, then step 3..." ✅ "Solve this problem" (let the model reason internally) ## Pattern Selection Guide | Task Type | Recommended Pattern | |-----------|-------------------| | Simple Q&A | Zero-shot | | Format-sensitive output | Few-shot | | Complex reasoning (standard LLM) | Chain-of-thought | | Complex reasoning (o1/o3/R1) | Simple, clear prompt | | Multi-step with tools | ReAct | | Creative exploration | Tree of Thoughts | | Code generation | PAL + structured output | | Classification | Few-shot with examples | | Summarization | Template + constraints |