prompt-engineering

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Prompt Engineering for Agentic Systems

面向智能体系统的提示词工程

Overview

概述

Generate optimized prompts for agentic systems with clear rationale for technique selection.

Core Principles:

Match technique to task - Different agent types require different prompting approaches
Trade-offs matter - Always consider cost, latency, and accuracy when selecting techniques
Structure over verbosity - Well-organized prompts outperform long unstructured ones
Test and iterate - Verify prompts work before deploying to production

For broader agentic system design (choosing workflows vs agents, ACI/tool specifications, guardrails, multi-agent patterns), see the ai-engineering skill.

生成针对智能体系统的优化提示词，并说明技术选择的理由。

核心原则:

匹配技术与任务——不同类型的智能体需要不同的提示词方法
权衡利弊——选择技术时始终要考虑成本、延迟和准确性
结构优于冗长——组织良好的提示词效果优于冗长无结构的提示词
测试与迭代——部署到生产环境前需验证提示词的有效性

如需更全面的智能体系统设计（选择工作流vs智能体、ACI/工具规范、防护机制、多智能体模式），请查看ai-engineering skill。

When to Use

适用场景

Invoke this skill when:

Creating prompts for tool-using agents (ReAct pattern)
Designing prompts for planning or strategy agents
Building prompts for data processing or validation agents
Reducing hallucinations in fact-based tasks
Optimizing prompt performance or cost

在以下情况调用本技能:

为工具型智能体（ReAct模式）创建提示词
为规划或策略型智能体设计提示词
为数据处理或验证型智能体构建提示词
在事实类任务中减少幻觉
优化提示词的性能或成本

Common Scenarios

常见场景

Scenario 1: Tool-Using Agent (ReAct)

场景1：工具型智能体（ReAct）

Use when: Agent needs to reason and use tools autonomously

markdown

undefined

适用场景：智能体需要自主推理并使用工具

markdown

undefined

SYSTEM

You are a research agent. Your goal is to gather information and synthesize findings.

INSTRUCTIONS

Follow this pattern for each action: Thought: [what you want to do] Action: [tool name and parameters] Observation: [result from tool]

When you have enough information, provide a final summary.

Follow this pattern for each action: Thought: [what you want to do] Action: [tool name and parameters] Observation: [result from tool]

When you have enough information, provide a final summary.

AVAILABLE TOOLS

search(query): Search for information
read(url): Read a webpage
finish(summary): Complete the task

search(query): Search for information
read(url): Read a webpage
finish(summary): Complete the task

STOP CONDITION

Stop when you have answered the user's question or gathered sufficient information.

---

Stop when you have answered the user's question or gathered sufficient information.

---

Scenario 2: Planning Agent (Tree of Thoughts)

场景2：规划智能体（Tree of Thoughts）

Use when: Agent needs to explore multiple approaches before committing

markdown

undefined

适用场景：智能体需要在执行前探索多种方案

markdown

undefined

TASK

Design a migration strategy from monolith to microservices.

INSTRUCTIONS

Generate 3 different approaches:

Approach 1: [description]

Pros: [list]
Cons: [list]
Effort: [estimate]

Approach 2: [description]

Pros: [list]
Cons: [list]
Effort: [estimate]

Approach 3: [description]

Pros: [list]
Cons: [list]
Effort: [estimate]

Then select the best approach and explain your reasoning.

---

Generate 3 different approaches:

Approach 1: [description]

Pros: [list]
Cons: [list]
Effort: [estimate]

Approach 2: [description]

Pros: [list]
Cons: [list]
Effort: [estimate]

Approach 3: [description]

Pros: [list]
Cons: [list]
Effort: [estimate]

Then select the best approach and explain your reasoning.

---

Scenario 3: Data Validation (Few-Shot with Negative Examples)

场景3：数据验证（带负例的少样本提示）

Use when: Agent needs consistent output format and should avoid common errors

markdown

undefined

适用场景：智能体需要输出格式一致，且需避免常见错误

markdown

undefined

TASK

Validate email addresses and return structured JSON.

VALID EXAMPLES

Input: user@example.com Output: {"valid": true, "reason": "proper email format"}

Input: user.name@company.co.uk Output: {"valid": true, "reason": "proper email format"}

Input: user@example.com Output: {"valid": true, "reason": "proper email format"}

Input: user.name@company.co.uk Output: {"valid": true, "reason": "proper email format"}

INVALID EXAMPLES (what NOT to accept)

Input: user@.com Output: {"valid": false, "reason": "invalid domain format"}

Input: @example.com Output: {"valid": false, "reason": "missing local part"}

Input: user example.com Output: {"valid": false, "reason": "missing @ symbol"}

Input: user@.com Output: {"valid": false, "reason": "invalid domain format"}

Input: @example.com Output: {"valid": false, "reason": "missing local part"}

Input: user example.com Output: {"valid": false, "reason": "missing @ symbol"}

NOW VALIDATE

Input: {user_input}

---

Input: {user_input}

---

Scenario 4: Factual Accuracy (Chain-of-Verification)

场景4：事实准确性（验证链，CoVe）

Use when: Reducing hallucinations is critical

markdown

undefined

适用场景：减少幻觉至关重要

markdown

undefined

TASK

Explain how transformers handle long-context windows

STEP 1: Initial Answer

Provide your explanation...

STEP 2: Verification Questions

Generate 5 questions that would expose errors in your answer:

[question]
[question]
[question]
[question]
[question]

Generate 5 questions that would expose errors in your answer:

[question]
[question]
[question]
[question]
[question]

STEP 3: Answer Verification

Answer each verification question factually...

STEP 4: Final Answer

Refine your original answer based on verification results...

---

Refine your original answer based on verification results...

---

Scenario 5: Complex Decision (Structured Thinking)

场景5：复杂决策（结构化思维）

Use when: Agent needs to analyze trade-offs before deciding

markdown

undefined

适用场景：智能体需要在决策前分析权衡利弊

markdown

undefined

TASK

Recommend: microservices or monolith for our startup?

THINKING PROTOCOL

[UNDERSTAND]

Restate the problem in your own words
Identify what's actually being asked

[ANALYZE]

Break down into sub-components
Note assumptions and constraints

[STRATEGIZE]

Outline 2-3 approaches
Evaluate trade-offs

[EXECUTE]

Provide final recommendation
Explain reasoning

---

[UNDERSTAND]

Restate the problem in your own words
Identify what's actually being asked

[ANALYZE]

Break down into sub-components
Note assumptions and constraints

[STRATEGIZE]

Outline 2-3 approaches
Evaluate trade-offs

[EXECUTE]

Provide final recommendation
Explain reasoning

---

Scenario 6: Self-Improving Output (Self-Refine)

场景6：自我优化输出（自我修正）

Use when: You want the agent to review and improve its own work

markdown

undefined

适用场景：希望智能体审核并改进自身输出

markdown

undefined

TASK

Write a README for the checkout API

STEP 1: Initial Draft

[Generate initial README]

STEP 2: Critique

Identify 3-5 improvements needed:

[weakness 1]
[weakness 2]
[weakness 3]

Identify 3-5 improvements needed:

[weakness 1]
[weakness 2]
[weakness 3]

STEP 3: Refinement

Rewrite addressing all identified improvements...

---

Rewrite addressing all identified improvements...

---

Quick Decision Tree

快速决策树

Use this table to select techniques quickly:

Agent Characteristic	Recommended Technique
Uses tools autonomously	ReAct
Planning/strategy with alternatives	Tree of Thoughts
High-stakes correctness	Self-Consistency
Factual accuracy, hallucination reduction	Chain-of-Verification (CoVe)
Single-path complex reasoning	Chain of Thought
Complex decisions with trade-offs	Structured Thinking Protocol
Reducing bias, multiple viewpoints	Multi-Perspective Prompting
Uncertainty quantification	Confidence-Weighted Prompting
Proprietary documentation, prevent hallucinations	Context Injection with Boundaries
Self-review and improvement	Self-Refine
Breaking complex problems into subproblems	Least-to-Most Prompting
High-quality content through multiple passes	Iterative Refinement Loop
Multi-stage workflows with specialized prompts	Prompt Chaining
Improving recall for factual questions	Generated Knowledge Prompting
Unclear how to structure the prompt	Meta-Prompting (nuclear option)
Strict technical requirements	Constraint-First Prompting
Requires consistent format/tone	Few-Shot (supports negative examples)
Simple, well-defined task	Zero-Shot
Domain-specific expertise	Role Prompting
Procedural workflow	Instruction Tuning

For detailed decision logic with branching, see decision-tree.md

使用下表快速选择技术:

智能体特征	推荐技术
自主使用工具	ReAct
需规划/策略与多方案	Tree of Thoughts
高风险场景下的正确性	Self-Consistency
事实准确性、减少幻觉	Chain-of-Verification (CoVe)
单路径复杂推理	Chain of Thought
需权衡利弊的复杂决策	结构化思维协议
减少偏见、多视角	多视角提示法
不确定性量化	置信度加权提示法
专有文档、防止幻觉	带边界的上下文注入
自我审核与改进	Self-Refine
将复杂问题拆解为子问题	由易到难提示法
多轮生成高质量内容	迭代优化循环
多阶段工作流与专用提示词	提示词链式调用
提升事实问题的召回率	生成式知识提示法
不清楚如何构建提示词	元提示法（终极方案）
严格技术要求	约束优先提示法
需一致格式/语气	少样本提示（支持负例）
简单、定义明确的任务	零样本提示
领域特定专业知识	角色提示法
流程化工作流	指令微调

如需带分支的详细决策逻辑，请查看 decision-tree.md

Technique Reference

技术参考

All available techniques with examples, use cases, and risks: techniques.md

所有可用技术的示例、适用场景与风险：techniques.md

Critical Anti-Patterns

关键反模式

Common mistakes to avoid: anti-patterns.md

Critical warnings:

Do NOT use ReAct without tools - Adds unnecessary complexity
Do NOT use Tree of Thoughts for deterministic problems - Single correct answer doesn't need alternatives
Do NOT use vague roles - "Expert" without scope provides little benefit
Do NOT omit stop conditions - Agents may continue indefinitely
Do NOT use Self-Refine for objective tasks - Calculations don't need self-critique

需避免的常见错误：anti-patterns.md

关键警告:

请勿在无工具的情况下使用ReAct——会增加不必要的复杂度
请勿在确定性问题中使用Tree of Thoughts——有唯一正确答案时无需多方案
请勿使用模糊角色——无明确范围的“专家”几乎没有价值
请勿省略停止条件——智能体可能无限运行
请勿在客观任务中使用Self-Refine——计算类任务不需要自我批判

Canonical Template

标准模板

Use this template as the foundation for generated prompts: template.md

Basic structure:

markdown

undefined

使用本模板作为生成提示词的基础：template.md

基本结构:

markdown

undefined

SYSTEM / ROLE

You are a [specific role] with authority over [explicit scope] Boundaries: [what you must NOT do]

TASK

[Single clear goal in one sentence]

INSTRUCTIONS

Follow these steps:

[First step]
[Second step]

Constraints:

[Specific limits]
[Format requirements]

Follow these steps:

[First step]
[Second step]

Constraints:

[Specific limits]
[Format requirements]

STOP CONDITION

Stop when: [success criteria]

undefined

Stop when: [success criteria]

undefined

Output Rationale Template

输出理由模板

When delivering a generated prompt, use this structure:

markdown

undefined

交付生成的提示词时，请使用以下结构:

markdown

undefined

Generated Prompt for [Agent Name/Type]

[prompt in code block]

Rationale

Agent Type: [Tool-using / Planner / Conversational / Data-processor]

Task Complexity: [Simple / Multi-step / Planning-heavy]

Techniques Used:

[Technique]: [Why it works for this use case]

Expected Behavior: [What the agent will do]

Trade-offs: [Cost, latency, flexibility - ALWAYS include if technique increases tokens or latency]

Considerations: [Edge cases, limitations, or risks]

undefined

Agent Type: [Tool-using / Planner / Conversational / Data-processor]

Task Complexity: [Simple / Multi-step / Planning-heavy]

Techniques Used:

[Technique]: [Why it works for this use case]

Expected Behavior: [What the agent will do]

Trade-offs: [Cost, latency, flexibility - ALWAYS include if technique increases tokens or latency]

Considerations: [Edge cases, limitations, or risks]

undefined

Guardrail Rule

防护规则

If a prompt increases latency, token usage, or operational cost, this MUST be stated explicitly in the rationale under "Trade-offs."

Techniques that increase cost/latency:

Self-Consistency (multiple generations)
Chain-of-Verification (multiple passes)
Iterative Refinement Loop (multiple passes)
Self-Refine (multiple passes)
Tree of Thoughts (exploring alternatives)
Least-to-Most Prompting (sequential subproblems)

如果提示词会增加延迟、token使用量或运营成本，必须在理由的“权衡利弊”部分明确说明。

会增加成本/延迟的技术:

Self-Consistency（多轮生成）
Chain-of-Verification（多轮验证）
迭代优化循环（多轮生成）
Self-Refine（多轮修正）
Tree of Thoughts（探索多方案）
由易到难提示法（顺序处理子问题）