prompting-pattern-library（提示模式库）

Prompting Pattern Library

Version 1.0 | October 2025 | Tested with Claude 3.5/4, GPT-4/4o, Gemini 1.5 Pro

Navigation

📖 Full Documentation: See README.md for complete navigation, use case index, and version notes.

Quick access by need:

• Learning prompting: Start here, then read Prompt Patterns
• Debugging prompts: Jump to Failure Modes
• Building agents: See Orchestration Patterns
• Model optimization: Review Model Quirks

Overview

This skill provides a comprehensive library of prompting patterns, anti-patterns, and model-specific guidance for effective LLM interactions. Use this when creating educational content about prompting, analyzing prompt quality, or explaining prompting techniques to technical and non-technical audiences.

What's included:

• 25+ proven prompting patterns with "why it works" analysis
• Common failure modes with diagnosis and fixes
• Model-specific guidance (Claude, GPT-4, Gemini)
• Advanced orchestration patterns for agent systems
• Cross-references throughout for deep-dive learning

Quick Reference: Common Prompting Patterns

Structural Patterns

Role Prompting: Assign a specific role or persona to frame the response Chain-of-Thought (CoT): Request step-by-step reasoning before final answer Few-Shot Learning: Provide examples of desired input-output pairs Zero-Shot with Instructions: Detailed task description without examples Tree of Thoughts: Explore multiple reasoning paths before choosing best

Output Control Patterns

Structured Output: Request specific formats (JSON, XML, tables, lists) Delimiters: Use clear separators for inputs, examples, and instructions Length Control: Specify desired output length explicitly Style Constraints: Define tone, formality, audience level

Reasoning Enhancement Patterns

Self-Consistency: Generate multiple solutions and select most common Reflection: Ask model to critique its own output Decomposition: Break complex tasks into smaller sub-tasks Analogical Reasoning: Request analogies or comparisons

Retrieval Patterns

Citation Requirements: Demand sources and evidence Fact-Checking: Request verification of claims Knowledge Boundaries: Ask model to acknowledge uncertainty

See references/prompt-patterns.md for comprehensive pattern catalog with examples.

When to Read References

Always Read First

Creating prompting educational content: Read references/prompt-patterns.md for pattern catalog with "why it works" analysis Debugging problematic prompts: Read references/failure-modes.md for common issues and fixes with cross-referenced solutions Cross-model implementation: Read references/model-quirks.md for model-specific considerations and optimization Building agent systems: Read references/orchestration-patterns.md for multi-step workflows and agentic architectures

Read When Needed

Advanced pattern implementation: Review specific patterns in references for detailed guidance and research basis Teaching prompting: Use examples from references as teaching materials with "why it works" explanations Optimizing existing prompts: Consult failure modes to identify weaknesses, then apply patterns from prompt-patterns.md Agent orchestration: Reference orchestration-patterns.md for planner-executor, multi-agent collaboration, and evaluation loops

Core Principles for Effective Prompting

Specificity Over Generality

Vague: "Write about AI" Specific: "Write a 500-word technical explanation of transformer attention mechanisms for software engineers with no ML background"

Provide Context Explicitly

Poor context: "Fix this code" Good context: "Fix this Python function that should validate email addresses. Current issue: it fails on addresses with plus signs. Python 3.11, standard library only."

Use Examples When Precision Matters

For tasks requiring specific formats or styles, provide 2-3 high-quality examples rather than lengthy descriptions. Examples communicate requirements more precisely than instructions alone.

Structure Complex Prompts

For multi-part tasks, use clear sections:

1. Context and background
2. Specific task requirements
3. Output format specifications
4. Constraints and limitations
5. Examples (if applicable)

Iterate Based on Output

Prompting is experimental. Start simple, observe failure modes, refine incrementally. Most effective prompts emerge through iteration, not perfect first attempts.

Model-Specific Considerations

Different models respond differently to identical prompts. Key differences:

Claude (Anthropic): Strong with structured output, detailed reasoning, and nuanced tasks. Responds well to polite, conversational prompts. Excellent at maintaining context over long conversations.

GPT-4 (OpenAI): Versatile across domains, strong creative writing, good instruction-following. Benefits from explicit structure. Can be more prone to confident errors.

Gemini (Google): Strong multimodal capabilities, good at analytical tasks. May require more explicit formatting instructions.

See references/model-quirks.md for detailed model-specific patterns and anti-patterns.

Common Failure Modes

The "Too Polite" Problem

Over-apologetic prompts waste tokens and can reduce output quality. Be direct and clear rather than excessively polite.

Implicit Assumptions

Models cannot read your mind. What seems obvious to you must be stated explicitly. Common implicit assumptions that cause failures:

• Desired output format
• Audience level
• Required depth of detail
• Constraints (time period, geography, etc.)

Conflicting Instructions

When instructions contradict each other, models exhibit unpredictable behavior. Example conflict: "Be concise but include comprehensive detail."

Ambiguous Success Criteria

"Make it better" is not actionable. Define what "better" means: faster, more accurate, more readable, more maintainable, etc.

See references/failure-modes.md for comprehensive failure patterns and fixes.

Using This Skill for Content Creation

Educational Content

When writing prompting guides or tutorials, use patterns from references as examples. Structure content to move from simple (zero-shot) to complex (chain-of-thought, tree-of-thoughts) patterns.

Prompt Analysis

To analyze prompt effectiveness, compare against patterns in references. Identify which patterns are present or absent, check for common failure modes.

Prompt Improvement

To improve existing prompts:

1. Identify the task type and check appropriate patterns in references
2. Check against failure modes in references/failure-modes.md
3. Apply relevant structural improvements
4. Test and iterate

Bundled References

references/prompt-patterns.md

Comprehensive catalog of 25+ prompting patterns with:

• Pattern name and description
• When to use each pattern
• Concrete examples with "why it works" analysis
• Variations and combinations
• Common pitfalls
• Cross-references to failure modes and model quirks
• Research basis where applicable

references/failure-modes.md

Common prompting failures organized by:

• Failure type (ambiguity, contradiction, assumption, etc.)
• Symptoms and diagnosis
• Root cause analysis
• Fixes with examples
• Cross-references to relevant patterns
• Prevention checklist

references/model-quirks.md

Model-specific guidance covering:

• Claude-specific patterns and anti-patterns
• GPT-specific patterns and anti-patterns
• Gemini-specific patterns and anti-patterns
• Cross-model considerations
• When model choice matters
• Model selection decision tree
• Testing strategies across models

references/orchestration-patterns.md

Advanced patterns for multi-step AI workflows:

• Meta-prompting patterns (prompt generation, optimization loops)
• Multi-agent orchestration (planner-executor, specialist collaboration)
• Evaluation and refinement loops (generate-critique-revise, ensemble evaluation)
• Tool integration patterns (tool selection, function call orchestration)
• Memory and state management (stateful conversations, context summarization)
• Pattern composition for production systems
• Performance optimization strategies