--- name: prompt-engineering description: >- Craft, debug, and improve prompts, skills, agents, and output styles. Use when working with any AI instructions. --- # Prompt Engineering Design effective prompts for LLMs. Apply when crafting skills, agents, output styles, system prompts, or any AI instructions. ## Quick Start Most prompts need only three things: 1. **Clear task** — action verb + objective 2. **Output format** — explicit structure 3. **One example** — if format matters Start here. Add complexity only when this fails. ``` Summarize the following article in 3 bullet points. Each bullet should be under 20 words. Focus on business implications. Example output: - Revenue increased 15% due to new product line launch - Operating costs reduced through automation initiatives - Market share expanded in European regions Article: {{ARTICLE}} ``` ## What's Wrong With Your Prompt? | Symptom | Fix | Details | |---------|-----|---------| | Wrong format | Add explicit format + example | [below](#output-format) | | Missing information | Be more specific about what to include | [below](#be-specific) | | Hallucination | Add context, request citations | [below](#provide-context) | | Ignores instructions | Move instructions to end, use XML tags | [long-context.md](references\/long-context.md) | | Complex reasoning fails | Add "Think step by step" | [reasoning-techniques.md](references\/reasoning-techniques.md) | | Inconsistent results | Add 3-5 examples | [learning-paradigms.md](references\/learning-paradigms.md) | | Too verbose | Specify word\/sentence limits | [below](#be-specific) | | Security concerns | Validate input, filter output | [security.md](references\/security.md) | ## Deep Dive References | Topic | When to Read | |-------|--------------| | [reasoning-techniques.md](references\/reasoning-techniques.md) | Multi-step reasoning, math, logic, CoT variants | | [learning-paradigms.md](references\/learning-paradigms.md) | Zero\/few-shot design, example selection | | [workflow-patterns.md](references\/workflow-patterns.md) | Multi-prompt pipelines, iterative refinement | | [security.md](references\/security.md) | Untrusted input, prompt injection defense | | [optimization-strategies.md](references\/optimization-strategies.md) | Prompting vs RAG vs fine-tuning | | [claude-specific.md](references\/claude-specific.md) | Prefilling, extended thinking, system prompts | | [long-context.md](references\/long-context.md) | 20K+ token documents | | [agent-patterns.md](references\/agent-patterns.md) | Tool use, ReAct, PAL patterns | --- ## Core Techniques Use these in order — most problems are solved by #1-3. ### 1. Be Clear and Direct **The golden rule:** Show your prompt to a colleague with minimal context. If they're confused, Claude will be too. #### Provide Context Tell Claude: - What the task results will be used for - Who the audience is - What success looks like #### Be Specific | Vague | Specific | |-------|----------| | "Summarize this" | "Summarize in 3 bullets, each under 20 words" | | "Make it better" | "Fix grammar errors, reduce word count by 30%" | | "Analyze the data" | "Calculate YoY growth, identify top 3 trends" | #### Output Format Always specify format explicitly. Show an example if structure matters: ``` Extract the following as JSON: - Product name - Price (number only) - In stock (boolean) Example output: {"name": "Widget Pro", "price": 29.99, "in_stock": true} ``` ### 2. Use Examples (Few-Shot) 3-5 examples typically sufficient. Cover edge cases. ``` Text: "Great service!" → Positive Text: "Worst purchase ever" → Negative Text: "It works as expected" → Neutral Text: "Absolutely love it!" → ? ``` **Tips:** - Order: simple → complex - Diverse, representative examples - Wrap in `` tags for clarity See [learning-paradigms.md](references\/learning-paradigms.md) for example selection strategies. ### 3. Use XML Tags Separate components for clarity and parseability: ```xml Analyze the contract for risks. <\/instructions> {{CONTRACT_TEXT}} <\/contract> List risks in tags, recommendations in . <\/output_format> ``` **Best practices:** - Consistent tag names throughout - Reference tags in instructions: "Using the contract in ..." - Nest for hierarchy: `...<\/inner><\/outer>` ### 4. Let Claude Think (Chain-of-Thought) For complex reasoning, ask Claude to show its work: ``` Think through this step by step, then provide your answer. ``` Or use structured tags: ``` Think through this in tags. Then provide your answer in tags. ``` **Critical:** Claude must output its thinking. Without outputting the thought process, no thinking actually occurs. See [reasoning-techniques.md](references\/reasoning-techniques.md) for CoT variants, Tree-of-Thoughts, and extended thinking. ### 5. Use Sequential Steps For multi-step tasks, number the steps: ``` Your task is to anonymize customer feedback. Instructions: 1. Replace customer names with "CUSTOMER_[ID]" 2. Replace emails with "EMAIL_[ID]@example.com" 3. Redact phone numbers as "PHONE_[ID]" 4. Leave product names intact 5. Output only processed messages, separated by "---" ``` Numbered steps ensure Claude follows exact sequence. --- ## Choosing a Technique ``` Simple task, clear format → Zero-shot (just ask) Need consistent format → Few-shot (3-5 examples) Complex reasoning → Chain-of-Thought Very complex problem → Extended Thinking Multi-step workflow → Prompt Chaining Need external information → ReAct (reasoning + tool use) Precise calculation → PAL (generate code) ``` For advanced patterns, see [agent-patterns.md](references\/agent-patterns.md). --- ## Claude-Specific Features ### Prefilling Start Claude's response to control format: ```python messages=[ {"role": "user", "content": "Extract as JSON: ..."}, {"role": "assistant", "content": "{"} # Prefill ] ``` Claude continues from `{`, outputting pure JSON without preamble. ### System Prompts (Roles) Use roles for domain expertise: ```python system="You are a senior security auditor specializing in web applications." ``` Role prompting activates domain-specific knowledge and reasoning patterns. ### Extended Thinking For very complex problems, use extended thinking mode (API parameter) instead of standard CoT. Use high-level guidance, not prescriptive steps. See [claude-specific.md](references\/claude-specific.md) for details. --- ## Quality Checklist Before finalizing a prompt: - [ ] Task is clear (single action verb) - [ ] Output format is explicit - [ ] Constraints are specific (not vague) - [ ] Examples cover edge cases (if using few-shot) - [ ] Golden rule passed (colleague wouldn't be confused) - [ ] Security considered (if handling untrusted input)