prompt-engineering（及时工程）

Prompt Engineering

Overview

This skill helps create highly effective prompts by selecting the optimal technique and format based on task characteristics. Analyzes complexity, target LLM, accuracy needs, and context to recommend the best approach from 10+ proven methods and 4 structured formats.

Quick Start Decision Tree

Answer these questions to find the right approach:

1. Task Complexity?

• Simple → Zero-Shot (see references/zero-shot.md)
• Need format/style → Few-Shot (see references/few-shot.md)
• Multi-step reasoning → Chain of Thought (see references/chain-of-thought.md)
• Tool/API use → ReAct (see references/react.md)
• Complex planning → Tree of Thoughts (see references/tree-of-thoughts.md)
• High-stakes → Self-Consistency (see references/self-consistency.md)

2. Target LLM?

• Claude → XML format (see references/xml-format.md)
• GPT → JSON format (see references/json-format.md)
• Multi-LLM → JSON (portable)
• Human-editable → YAML (see references/yaml-format.md)

3. Output Use?

• Code/API → JSON
• Complex hierarchy → XML (if Claude) or JSON
• Human editing → YAML or Natural Language
• Simple explanation → Natural Language (see references/natural-language.md)

For detailed decision matrix: references/decision_matrix.md

Method Selection Quick Reference

Format Selection Quick Reference

Common Patterns At-A-Glance

Zero-Shot Template

Task: [X]
Requirements: [Y]
Output: [Z]

Few-Shot Template

Task: [X]
Examples:
- Input: A → Output: B
- Input: C → Output: D
Your turn: Input: E → Output: ?

Chain of Thought Template

Problem: [X]
Let's think step by step:
1. [Step 1]
2. [Step 2]
3. [Step 3]
Answer: [Y]

ReAct Template

Thought: [Reasoning]
Action: [Tool/action]
Observation: [Result]
[Repeat]

For complete templates and examples, see individual method references.

Common Pitfalls & Quick Fixes

❌ Ambiguous Instructions

Bad: "Make this better" Good: "Improve by: 1) Add error handling, 2) Optimize to O(n log n), 3) Add docs"

❌ Wrong Format for LLM

Bad: JSON for Claude complex hierarchy Good: XML for Claude, JSON for GPT/APIs

❌ No Examples When Needed

Bad: "Extract features in structured format" Good: Show 2-3 concrete input→output examples

❌ Overcomplicating Simple Tasks

Bad: Tree of Thoughts for "Convert 25°C to F" Good: Simple zero-shot instruction

For complete pitfalls guide: references/pitfalls.md

Advanced Techniques

Combining Methods

• Few-Shot + CoT: Examples with reasoning steps (see references/advanced-combinations.md)
• ReAct + Self-Consistency: Multiple tool paths, compare results
• ToT + XML: Claude-optimized complex planning

Meta-Prompting

Ask LLM to help design the prompt:

I need a prompt for [task].
Task characteristics: [complexity, LLM, output use, accuracy needs]
Recommend: 1) Technique, 2) Format, 3) Draft template, 4) Why

Prompt Chaining

Break complex tasks into sequential prompts. See: references/prompt-chaining.md

Token Efficiency Tips

✓ More Efficient          ✗ Less Efficient
- Zero-shot for simple    - 10+ examples
- Concise instructions    - Verbose repetition
- JSON for API parsing    - XML for API parsing
- Direct examples         - Over-explained examples

Typical token counts:

• Zero-Shot: 50-200 tokens
• Few-Shot (3 examples): 200-800 tokens
• Chain of Thought: 100-500 tokens
• ReAct: 300-1000 tokens
• Tree of Thoughts: 500-2000+ tokens
• Self-Consistency: 500-3000+ tokens

Implementation Checklist

When creating a prompt:

• Identify task complexity → Choose method
• Identify target LLM → Choose format
• Write clear, specific instructions
• Add examples if Few-Shot/Few-Shot CoT
• Specify output format explicitly
• Include constraints and requirements
• Test with sample inputs
• Validate outputs
• Refine based on results

Resources

References (Detailed Guides)

• decision_matrix.md - Comprehensive decision framework
• zero-shot.md - Zero-shot prompting guide
• few-shot.md - Few-shot prompting guide
• chain-of-thought.md - CoT prompting guide
• react.md - ReAct prompting guide
• tree-of-thoughts.md - ToT prompting guide
• self-consistency.md - Self-consistency guide
• xml-format.md - XML format best practices
• json-format.md - JSON format best practices
• yaml-format.md - YAML format best practices
• natural-language.md - Natural language best practices
• pitfalls.md - Common mistakes and solutions
• advanced-combinations.md - Combining techniques
• prompt-chaining.md - Sequential prompt patterns

Assets (Templates)

• templates/ - Ready-to-use templates for common scenarios (coming soon)

Navigate to specific references above for detailed implementation guides, templates, and examples.