技能 prompt-engineering-patterns

📝

prompt-engineering-patterns

Name: prompt-engineering-patterns
Author: wshobson

安全 ⚙️ 外部命令📁 檔案系統存取🌐 網路存取

掌握提示工程以獲得更好的AI結果

大型語言模型在處理提示不佳的提示時會產生不一致的結果。本技能提供了經過實際驗證的模式和模板，用於思維鏈推理、少量學習和系統性提示優化，以提高輸出品質。

支援: Claude Codex Code(CC)

🥈 81 白銀

下載技能 ZIP

在 Claude 中上傳

前往設定 → 功能 → 技能 → 上傳技能

開啟並開始使用

測試它

正在使用「prompt-engineering-patterns」。撰寫一個提示來摘要客戶回饋

預期結果:

從系統角色開始：您是專業分析師。
新增特定約束：以3個要點摘要。
包含範例：顯示回饋類別的輸入輸出配對。
定義格式：為每個摘要使用一致的結構。

安全審計

安全

v4 • 1/17/2026

This is a documentation-focused skill containing markdown guides and a local Python utility script for prompt optimization. The 228 static findings are false positives triggered by documentation patterns: backticks in Python code examples misinterpreted as shell commands, cryptographic terminology (SHA, MD5) mentioned in text, and references to API keys and file paths. The skill makes no network calls, has no sensitive filesystem access, and does not execute external commands. The optimize-prompt.py script uses a mock LLM client for local testing only.

已掃描檔案

2,919

分析行數

發現項

審計總數

風險因素

⚙️ 外部命令 (169)

📁 檔案系統存取 (3)

references/prompt-optimization.md:128 references/prompt-optimization.md:128 scripts/optimize-prompt.py:229

🌐 網路存取 (1)

skill-report.json:6

審計者: claude 查看審計歷史 →

品質評分

架構

100

可維護性

內容

社群

100

安全

規範符合性

你能建構什麼

優化生產提示

使用A/B測試框架系統性地測試和改進生產LLM應用程式的提示。

建立模板庫

建立具有變數插值的可重複使用提示模板，以實現一致的內容生成。

應用進階技術

為複雜推理任務實施思維鏈和自我一致性模式。

試試這些提示

簡單分類

將此文字分類為以下類別之一：正面、負面、中立。

文字：{text}

類別：

少量學習萃取

以JSON格式萃取資訊。

範例：
文字：Apple執行長Tim Cook宣布新手機。
輸出：{"persons":["Tim Cook"],"organizations":["Apple"],"products":["iPhone"]}

文字：{text}

輸出：

思維鏈

逐步解決此問題。

問題：{problem}

步驟1：識別我們已知的信息
步驟2：確定方法
步驟3：計算
步驟4：驗證

答案：

自我一致性

用三種不同的方法解決此問題。然後找出出現最頻繁的答案。

問題：{problem}

方法1：
結果：

方法2：
結果：

方法3：
結果：

最終答案（最常見）：

最佳實務

Be specific about format, length, and style requirements rather than relying on implied instructions
Use few-shot examples to demonstrate the exact output format you need, especially for structured data
Test prompts on edge cases and diverse inputs before deploying to production

避免

Overloading prompts with too many examples, causing token limits to reduce space for actual input
Using vague instructions like 'be helpful' or 'be accurate' that different models interpret differently
Skipping verification steps for factual or logical outputs that require validation

常見問題

Which LLMs work with these patterns?

Patterns work with Claude, GPT-4, Claude Code, and most instruction-tuned models. Some techniques like chain-of-thought work best on reasoning-capable models.

What is the optimal number of few-shot examples?

Most tasks perform well with 3 to 5 examples. More examples can dilute focus and consume token budget. Test different counts for your specific use case.

How do I integrate with my existing codebase?

The skill provides template systems and Python utilities. Adapt the PromptTemplate classes to your LLM client. The optimize-prompt.py script shows a testing workflow.

Is my data sent anywhere?

No. This skill runs locally. The reference materials and utility scripts operate entirely on your machine. No external network calls are made by any component.

Why do my prompts work differently across models?

Models have different training and capabilities. Test and adjust templates per model. Chain-of-thought works better on reasoning models. Some models need more explicit format instructions.

How does this compare to other prompting skills?

This skill focuses on production-ready patterns with systematic optimization workflows. It covers template systems, A/B testing, and evaluation metrics for real-world deployment.