prompt-engineering

Source Cursor rule: `.cursor/rules/prompt-engineering.mdc`. Original file scope: `.cursor/rules/*.mdc`. Original Cursor alwaysApply: `false`. # Prompt Engineering Best Practices Based on [Claude's Prompt Engineering Documentation](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/overview) ## Core Principles ### 1. Define Success Criteria First Before writing prompts: - **Establish clear objectives**: What constitutes a successful response? - **Create evaluation metrics**: How will you measure prompt effectiveness? - **Draft and iterate**: Start with a first draft and refine based on results ### 2. When to Use Prompt Engineering vs Fine-tuning Prompt engineering is preferred because: - **Resource efficient**: Only requires text input, no GPUs - **Cost effective**: Uses base model pricing - **Maintains updates**: Works across model versions - **Time saving**: Instant results vs hours/days for fine-tuning - **Minimal data needs**: Works with zero-shot or few-shot - **Flexible iteration**: Quick experimentation cycle - **Preserves knowledge**: No catastrophic forgetting - **Transparent**: Human-readable, easy to debug --- ## The 6 Core Techniques ### 1. Be Clear and Direct **Principle**: Provide explicit, unambiguous instructions. ``` ❌ Bad: "Tell me about it" ✅ Good: "Summarize the following article in three bullet points, focusing on key findings" ❌ Bad: "Help with code" ✅ Good: "Debug this Python function that should return the sum of even numbers in a list" ``` **Tips**: - State the task explicitly at the start - Specify the desired output format (bullet points, JSON, paragraphs) - Include constraints (word count, tone, audience) - Mention what to include AND what to exclude ### 2. Use Examples (Multishot Prompting) **Principle**: Show the model what you want through examples. ```xml <examples> <example> <input>The movie was absolutely terrible, waste of time</input> <output>{"sentiment": "negative", "confidence": 0.95}</output> </example> <example> <input>Decent film, not great but watchable</input> <output>{"sentiment": "neutral", "confidence": 0.7}</output> </example> <example> <input>Best movie I've seen this year!</input> <output>{"sentiment": "positive", "confidence": 0.9}</output> </example> </examples> Now analyze: "The special effects were amazing but the plot was confusing" ``` **Tips**: - Include 3-5 diverse examples covering edge cases - Show examples of BOTH good and bad outputs - Match example complexity to your actual use case - Order examples from simple to complex ### 3. Let Claude Think (Chain of Thought) **Principle**: Encourage step-by-step reasoning for complex tasks. ```xml <instruction> Solve this problem step by step. Show your reasoning before giving the final answer. </instruction> <problem> A train leaves Station A at 9:00 AM traveling at 60 mph. Another train leaves Station B at 10:00 AM traveling at 80 mph toward Station A. The stations are 280 miles apart. When will the trains meet? </problem> <thinking> [Let Claude work through the problem here] </thinking> <answer> [Final answer after reasoning] </answer> ``` **Tips**: - Use phrases like "Think step by step" or "Explain your reasoning" - For complex tasks, explicitly request a thinking section - Chain of thought improves accuracy on math, logic, and multi-step problems - Can use `<thinking>` tags to separate reasoning from output ### 4. Use XML Tags **Principle**: Structure prompts with clear delimiters for better parsing. ```xml <context> You are helping debug a penetration testing tool that automates security scans. </context> <task> Analyze the following error log and identify the root cause. </task> <error_log> [2024-01-15 10:23:45] ERROR: Connection timeout after 30s [2024-01-15 10:23:45] DEBUG: Target: 192.168.1.1:443 [2024-01-15 10:23:45] DEBUG: Retry attempt 3 of 3 </error_log> <output_format> Provide your analysis in this format: - Root cause: [one sentence] - Evidence: [relevant log lines] - Recommended fix: [actionable steps] </output_format> ``` **Common XML Tags**: - `<context>` - Background information - `<task>` or `<instruction>` - What to do - `<examples>` - Sample inputs/outputs - `<constraints>` - Limitations or rules - `<output_format>` - Expected response structure - `<thinking>` - Reasoning section - `<answer>` - Final response ### 5. Give Claude a Role (System Prompts) **Principle**: Assign a persona to influence response style and expertise. ```xml <role> You are a senior security researcher with 15 years of experience in penetration testing. You specialize in web application security and have discovered multiple CVEs. You communicate findings clearly and prioritize actionable recommendations. </role> <task> Review this HTTP response and identify potential security vulnerabilities. </task> ``` **Effective Role Elements**: - Expertise level (senior, expert, specialist) - Domain knowledge (security, finance, medicine) - Communication style (technical, friendly, formal) - Priorities (accuracy, brevity, thoroughness) ### 6. Prefill Claude's Response **Principle**: Start the response to guide format and direction. ``` Human: List the top 3 security vulnerabilities in this code. ```