Appendix A: Advanced Prompting Techniques

附录 A：高级提示技术

Introduction to Prompting

提示技术简介

Prompting, the primary interface for interacting with language models, is the process of crafting inputs to guide the model towards generating a desired output. This involves structuring requests, providing relevant context, specifying the output format, and demonstrating expected response types. Well-designed prompts can maximize the potential of language models, resulting in accurate, relevant, and creative responses. In contrast, poorly designed prompts can lead to ambiguous, irrelevant, or erroneous outputs.

提示技术是与语言模型交互的核心方式，指通过精心设计输入来引导模型生成期望输出的过程。这包括构建请求结构、提供相关上下文、指定输出格式以及展示预期的响应模式。设计精良的提示能够充分发挥语言模型的潜力，产生准确、相关且富有创造性的响应。反之，设计不当的提示则可能导致模糊、不相关甚至错误的输出。

The objective of prompt engineering is to consistently elicit high-quality responses from language models. This requires understanding the capabilities and limitations of the models and effectively communicating intended goals. It involves developing expertise in communicating with AI by learning how to best instruct it.

提示工程的目标是从语言模型中持续获得高质量的输出。这需要深入理解模型的能力边界和局限性，并有效地传达预期目标。它涉及通过学习如何最优地指导 AI 来发展与 AI 沟通的专业技能。

This appendix details various prompting techniques that extend beyond basic interaction methods. It explores methodologies for structuring complex requests, enhancing the model’s reasoning abilities, controlling output formats, and integrating external information. These techniques are applicable to building a range of applications, from simple chatbots to complex multi-agent systems, and can improve the performance and reliability of agentic applications.

本附录详细介绍了超越基础交互方法的各类提示技术。它探讨了构建复杂请求、增强模型推理能力、控制输出格式以及整合外部信息的方法论。这些技术适用于构建从简单聊天机器人到复杂多智能体的各类应用，能够显著提升 Agentic 应用的性能和可靠性。 Agentic patterns, the architectural structures for building intelligent systems, are detailed in the main chapters. These patterns define how agents plan, utilize tools, manage memory, and collaborate. The efficacy of these agentic systems is contingent upon their ability to interact meaningfully with language models.

Agentic 模式是构建智能系统的架构蓝图，在主要章节中已有详细阐述。这些模式定义了智能体如何规划、使用工具、管理记忆和协同工作。这些 Agentic 系统的效能最终取决于它们与语言模型进行有意义交互的能力。

Core Prompting Principles

核心提示原则

Core Principles for Effective Prompting of Language Models:

与语言模型进行有效提示的核心原则：

Effective prompting rests on fundamental principles guiding communication with language models, applicable across various models and task complexities. Mastering these principles is essential for consistently generating useful and accurate responses.

有效的提示建立在与语言模型沟通的基本原则之上，这些原则适用于各种模型和任务复杂度。掌握这些原则对于持续生成有用且准确的响应至关重要。

Clarity and Specificity: Instructions should be unambiguous and precise. Language models interpret patterns; multiple interpretations may lead to unintended responses. Define the task, desired output format, and any limitations or requirements. Avoid vague language or assumptions. Inadequate prompts yield ambiguous and inaccurate responses, hindering meaningful output.

清晰性与具体性：指令应当明确无误且精确。语言模型基于模式进行推理；多重解释可能导致意外结果。需要明确定义任务、期望的输出格式以及任何限制或要求。避免使用模糊语言或做出假设。不充分的提示会产生模糊和不准确的响应，影响输出质量。

Conciseness: While specificity is crucial, it should not compromise conciseness. Instructions should be direct. Unnecessary wording or complex sentence structures can confuse the model or obscure the primary instruction. Prompts should be simple; what is confusing to the user is likely confusing to the model. Avoid intricate language and superfluous information. Use direct phrasing and active verbs to clearly delineate the desired action. Effective verbs include: Act, Analyze, Categorize, Classify, Contrast, Compare, Create, Describe, Define, Evaluate, Extract, Find, Generate, Identify, List, Measure, Organize, Parse, Pick, Predict, Provide, Rank, Recommend, Return, Retrieve, Rewrite, Select, Show, Sort, Summarize, Translate, Write.

简洁性：在确保具体性的同时，必须保持简洁。指令应当直截了当。不必要的措辞或复杂的句子结构可能混淆模型或掩盖核心指令。提示应当简单明了：对用户而言困惑的内容，对模型同样可能造成困惑。避免使用复杂语言和冗余信息。采用直接表达和主动动词来清晰界定期望操作。有效的动词包括：执行、分析、分类、归类、对比、比较、创建、描述、定义、评估、提取、查找、生成、识别、列出、测量、组织、解析、挑选、预测、提供、排序、推荐、返回、检索、重写、选择、显示、整理、总结、翻译、编写。

Using Verbs: Verb choice is a key prompting tool. Action verbs indicate the expected operation. Instead of “Think about summarizing this,” a direct instruction like “Summarize the following text” is more effective. Precise verbs guide the model to activate relevant training data and processes for that specific task.

善用动词：动词选择是提示工程的关键技巧。动作动词明确指示期望的操作。与其说”考虑总结这个”，不如直接使用”总结以下文本”这样的指令更为有效。精确的动词能够引导模型激活与特定任务相关的训练数据和流程。

Instructions Over Constraints: Positive instructions are generally more effective than negative constraints. Specifying the desired action is preferred to outlining what not to do. While constraints have their place for safety or strict formatting, excessive reliance can cause the model to focus on avoidance rather than the objective. Frame prompts to guide the model directly. Positive instructions align with human guidance preferences and reduce confusion.

指令优于约束：积极指令通常比消极约束更为有效。明确指定期望操作比罗列禁止事项效果更好。虽然约束在安全控制或严格格式要求中有其价值，但过度依赖可能使模型过度关注规避而非目标达成。构建提示时应直接引导模型。积极指令更符合人类指导习惯，并能减少混淆。

Experimentation and Iteration: Prompt engineering is an iterative process. Identifying the most effective prompt requires multiple attempts. Begin with a draft, test it, analyze the output, identify shortcomings, and refine the prompt. Model variations, configurations (like temperature or top-p), and slight phrasing changes can yield different results. Documenting attempts is vital for learning and improvement. Experimentation and iteration are necessary to achieve the desired performance.

实验与迭代：提示工程是一个迭代过程。找到最有效的提示需要多次尝试。从初步设计开始，进行测试，分析输出，识别不足，然后优化提示。模型变体、配置参数（如温度或 top-p）以及细微的措辞变化都可能产生不同结果。记录实验过程对于学习和改进至关重要。实验和迭代是实现预期性能的必要手段。

These principles form the foundation of effective communication with language models. By prioritizing clarity, conciseness, action verbs, positive instructions, and iteration, a robust framework is established for applying more advanced prompting techniques.

这些原则构成了与语言模型有效沟通的基础。通过优先考虑清晰性、简洁性、善用动词、积极指令和迭代流程，为应用更高级提示技术奠定了坚实基础。

Basic Prompting Techniques

基础提示技术

Building on core principles, foundational techniques provide language models with varying levels of information or examples to direct their responses. These methods serve as an initial phase in prompt engineering and are effective for a wide spectrum of applications.

基于核心原则，基础技术为语言模型提供不同层次的信息或示例来引导其响应。这些方法作为提示工程的入门阶段，适用于广泛的应用场景。

Zero-Shot Prompting

零样本提示（Zero-Shot）

Zero-shot prompting is the most basic form of prompting, where the language model is provided with an instruction and input data without any examples of the desired input-output pair. It relies entirely on the model’s pre-training to understand the task and generate a relevant response. Essentially, a zero-shot prompt consists of a task description and initial text to begin the process.

零样本提示是最基础的提示形式，语言模型仅接收指令和输入数据，不提供任何期望的输入-输出对示例。它完全依赖模型的预训练知识来理解任务并生成相关响应。本质上，零样本提示包含任务描述和启动过程的初始文本。

One-Shot Prompting

单样本提示（One-Shot）

One-shot prompting involves providing the language model with a single example of the input and the corresponding desired output prior to presenting the actual task. This method serves as an initial demonstration to illustrate the pattern the model is expected to replicate. The purpose is to equip the model with a concrete instance that it can use as a template to effectively execute the given task.

单样本提示在呈现实际任务前，向语言模型提供一个输入及其对应期望输出的示例。这种方法作为初始演示，展示模型应遵循的模式。目的是为模型提供具体实例，使其能够作为模板有效执行给定任务。

英文：’Please.’ 西班牙语：

Few-Shot Prompting

少样本提示（Few-Shot）

Few-shot prompting enhances one-shot prompting by supplying several examples, typically three to five, of input-output pairs. This aims to demonstrate a clearer pattern of expected responses, improving the likelihood that the model will replicate this pattern for new inputs. This method provides multiple examples to guide the model to follow a specific output pattern.

少样本提示在单样本基础上增强，提供多个（通常三到五个）输入-输出对示例。这旨在展示更清晰的预期响应模式，提高模型为新输入复制该模式的可能性。此方法通过多个示例引导模型遵循特定输出模式。

Review: “The acting was superb and the story was engaging.”
Sentiment: POSITIVE

评论：”表演精湛，故事引人入胜。” 情感：POSITIVE

Review: “It was okay, nothing special.”
Sentiment: NEUTRAL

评论：”还行，没什么特别的。” 情感：NEUTRAL

Review: “I found the plot confusing and the characters unlikable.”
Sentiment: NEGATIVE

评论：”我觉得情节混乱，角色不讨喜。” 情感：NEGATIVE

Review: “The visuals were stunning, but the dialogue was weak.”
Sentiment:

评论：”视觉效果惊艳，但对话薄弱。” 情感：

Understanding when to apply zero-shot, one-shot, and few-shot prompting techniques, and thoughtfully crafting and organizing examples, are essential for enhancing the effectiveness of agentic systems. These basic methods serve as the groundwork for various prompting strategies.

理解何时应用零样本、单样本和少样本提示技术，并精心设计和组织示例，对于提升 Agentic 系统的效能至关重要。这些基础方法为各种提示策略奠定了根基。

Structuring Prompts

提示结构设计

Beyond the basic techniques of providing examples, the way you structure your prompt plays a critical role in guiding the language model. Structuring involves using different sections or elements within the prompt to provide distinct types of information, such as instructions, context, or examples, in a clear and organized manner. This helps the model parse the prompt correctly and understand the specific role of each piece of text.

除了提供示例的基础技术外，提示的结构化方式在引导语言模型方面起着关键作用。结构化涉及在提示中使用不同部分或元素，以清晰有序的方式提供指令、上下文或示例等不同类型的信息。这有助于模型正确解析提示，理解每段文本的特定角色。

System Prompting

System prompting sets the overall context and purpose for a language model, defining its intended behavior for an interaction or session. This involves providing instructions or background information that establish rules, a persona, or overall behavior. Unlike specific user queries, a system prompt provides foundational guidelines for the model’s responses. It influences the model’s tone, style, and general approach throughout the interaction. For example, a system prompt can instruct the model to consistently respond concisely and helpfully or ensure responses are appropriate for a general audience. System prompts are also utilized for safety and toxicity control by including guidelines such as maintaining respectful language.

系统提示为语言模型设定整体上下文和目的，定义其在交互或会话中的预期行为。这涉及提供建立规则、角色或整体行为的指令或背景信息。与具体的用户查询不同，系统提示为模型的响应提供基础指导。它影响模型在整个交互过程中的语气、风格和总体方法。例如，系统提示可指示模型始终保持简洁有益的响应，或确保输出适合普通受众。系统提示还用于安全和内容控制，包含保持尊重语言等指导原则。

Furthermore, to maximize their effectiveness, system prompts can undergo automatic prompt optimization through LLM-based iterative refinement. Services like the Vertex AI Prompt Optimizer facilitate this by systematically improving prompts based on user-defined metrics and target data, ensuring the highest possible performance for a given task.

此外，为最大化其效果，系统提示可通过基于 LLM 的迭代改进进行自动优化。Vertex AI Prompt Optimizer 等服务通过根据用户定义的指标和目标数据系统性地优化提示来促进这一过程，确保特定任务的最佳性能。

Role Prompting

角色提示

Role prompting assigns a specific character, persona, or identity to the language model, often in conjunction with system or contextual prompting. This involves instructing the model to adopt the knowledge, tone, and communication style associated with that role. For example, prompts such as “Act as a travel guide” or “You are an expert data analyst” guide the model to reflect the perspective and expertise of that assigned role. Defining a role provides a framework for the tone, style, and focused expertise, aiming to enhance the quality and relevance of the output. The desired style within the role can also be specified, for instance, “a humorous and inspirational style.”

角色提示为语言模型分配特定角色、身份或专业背景，通常与系统或上下文提示结合使用。这涉及指示模型采用与该角色相关的知识、语气和沟通风格。例如，”扮演旅游指南”或”你是一位专业数据分析师”等提示引导模型体现所分配角色的视角和专长。定义角色为语气、风格和专业焦点提供框架，旨在提升输出的质量和相关性。还可指定角色内的期望风格，如”幽默且鼓舞人心的风格”。

Using Delimiters

使用分隔符

Effective prompting involves clear distinction of instructions, context, examples, and input for language models. Delimiters, such as triple backticks (\`\`\`), XML tags (\<instruction\>, \<context\>), or markers (—), can be utilized to visually and programmatically separate these sections. This practice, widely used in prompt engineering, minimizes misinterpretation by the model, ensuring clarity regarding the role of each part of the prompt.

有效的提示需要为语言模型清晰区分指令、上下文、示例和输入。可使用分隔符，如三重反引号（```）、XML 标签（、）或标记（---），在视觉和程序上分隔这些部分。这种在提示工程中广泛采用的做法，通过明确提示各部分的角色，最小化模型的误解。

Contextual Enginnering

上下文工程

Context engineering, unlike static system prompts, dynamically provides background information crucial for tasks and conversations. This ever-changing information helps models grasp nuances, recall past interactions, and integrate relevant details, leading to grounded responses and smoother exchanges. Examples include previous dialogue, relevant documents (as in Retrieval Augmented Generation), or specific operational parameters. For instance, when discussing a trip to Japan, one might ask for three family-friendly activities in Tokyo, leveraging the existing conversational context. In agentic systems, context engineering is fundamental to core agent behaviors like memory persistence, decision-making, and coordination across sub-tasks. Agents with dynamic contextual pipelines can sustain goals over time, adapt strategies, and collaborate seamlessly with other agents or tools—qualities essential for long-term autonomy. This methodology posits that the quality of a model’s output depends more on the richness of the provided context than on the model’s architecture. It signifies a significant evolution from traditional prompt engineering, which primarily focused on optimizing the phrasing of immediate user queries. Context engineering expands its scope to include multiple layers of information.

上下文工程与静态系统提示不同，它动态提供对任务和对话至关重要的背景信息。这种持续变化的信息帮助模型理解细微差别、回忆过往交互并整合相关细节，从而产生有根据的响应和更流畅的交流。示例包括先前对话、相关文档（如在检索增强生成中）或特定操作参数。例如，在讨论日本旅行时，可请求提供东京的三个适合家庭的活动，利用现有对话上下文。在 Agentic 系统中，上下文工程是核心智能体能力（如记忆持久性、决策制定和跨子任务协调）的基础。具备动态上下文管道的智能体能够维持目标、调整策略，并与其他智能体或人类协作——这些是实现长期自主性的关键特质。该方法论认为，模型输出的质量更多取决于所提供上下文的丰富度，而非模型架构本身。它标志着从传统提示工程的重大演进，传统提示工程主要聚焦于优化直接用户查询的措辞。上下文工程将其范畴扩展至包含多层信息。

These layers include:

这些层次包括：

The core principle is that even advanced models underperform with a limited or poorly constructed view of their operational environment. This practice reframes the task from merely answering a question to building a comprehensive operational picture for the agent. For example, a context-engineered agent would integrate a user’s calendar availability (tool output), the professional relationship with an email recipient (implicit data), and notes from previous meetings (retrieved documents) before responding to a query. This enables the model to generate highly relevant, personalized, and pragmatically useful outputs. The “engineering” aspect involves creating robust pipelines to fetch and transform this data at runtime and establishing feedback loops to continually improve context quality.

核心原则是，即使是先进模型，若对其操作环境的认知有限或构建不当，表现也会欠佳。这种做法将任务从单纯回答问题重新定义为为智能体构建全面的操作图景。例如，一个经过上下文工程设计的智能体在回复用户查询前，会整合用户的日历可用性（工具输出）、与邮件收件人的专业关系（隐式数据）以及过往会议记录（检索文档）。这使得模型能够生成高度相关、个性化且实用的输出。”工程”方面涉及构建稳健的管道以在运行时获取和转换这些数据，并建立反馈循环以持续提升上下文质量。

To implement this, specialized tuning systems, such as Google’s Vertex AI prompt optimizer, can automate the improvement process at scale. By systematically evaluating responses against sample inputs and predefined metrics, these tools can enhance model performance and adapt prompts and system instructions across different models without extensive manual rewriting. Providing an optimizer with sample prompts, system instructions, and a template allows it to programmatically refine contextual inputs, offering a structured method for implementing the necessary feedback loops for sophisticated Context Engineering.
This structured approach differentiates a rudimentary AI tool from a more sophisticated, contextually-aware system. It treats context as a primary component, emphasizing what the agent knows, when it knows it, and how it uses that information. This practice ensures the model has a well-rounded understanding of the user’s intent, history, and current environment. Ultimately, Context Engineering is a crucial methodology for transforming stateless chatbots into highly capable, situationally-aware systems.

为实现这一点，专门的调优系统（如 Google 的 Vertex AI prompt optimizer）可大规模自动化改进过程。通过基于样本输入和预定义指标系统评估响应，这些工具能提升模型性能，并在不同模型间调整提示和系统指令，无需大量手动重写。为优化器提供样本提示、系统指令和模板，使其能程序化地优化上下文输入，为实施复杂上下文工程所需的反馈循环提供结构化方法。

这种结构化方法将基础 AI 工具与更复杂的情境感知系统区分开来。它将上下文视为核心要素，强调智能体知晓什么、何时知晓以及如何运用这些信息。这种做法确保模型对用户的意图、历史和当前环境有全面理解。最终，上下文工程是将无状态聊天机器人转变为高能力情境感知系统的关键方法论。

Structured Output

结构化输出

Often, the goal of prompting is not just to get a free-form text response, but to extract or generate information in a specific, machine-readable format. Requesting structured output, such as JSON, XML, CSV, or Markdown tables, is a crucial structuring technique. By explicitly asking for the output in a particular format and potentially providing a schema or example of the desired structure, you guide the model to organize its response in a way that can be easily parsed and used by other parts of your agentic system or application. Returning JSON objects for data extraction is beneficial as it forces the model to create a structure and can limit hallucinations. Experimenting with output formats is recommended, especially for non-creative tasks like extracting or categorizing data.

通常，提示的目标不仅是获得自由形式的文本响应，而是以特定机器可读格式提取或生成信息。请求结构化输出（如 JSON、XML、CSV 或 Markdown 表格）是一项关键的结构化技术。通过明确要求特定格式输出并可能提供期望结构的模式或示例，您可以引导模型以易于被 Agentic 系统或其他应用组件解析和使用的方式组织响应。返回 JSON 对象进行数据提取的优势在于强制模型创建结构，从而限制幻觉产生。建议尝试不同输出格式，特别是对于数据提取或分类等非创意任务。

Effectively utilizing system prompts, role assignments, contextual information, delimiters, and structured output significantly enhances the clarity, control, and utility of interactions with language models, providing a strong foundation for developing reliable agentic systems. Requesting structured output is crucial for creating pipelines where the language model’s output serves as the input for subsequent system or processing steps.

Leveraging Pydantic for an Object-Oriented Facade: A powerful technique for enforcing structured output and enhancing interoperability is to use the LLM’s generated data to populate instances of Pydantic objects. Pydantic is a Python library for data validation and settings management using Python type annotations. By defining a Pydantic model, you create a clear and enforceable schema for your desired data structure. This approach effectively provides an object-oriented facade to the prompt’s output, transforming raw text or semi-structured data into validated, type-hinted Python objects.

有效利用系统提示、角色分配、上下文信息、分隔符和结构化输出，显著提升了与语言模型交互的清晰度、控制力和实用性，为构建可靠的 Agentic 系统奠定了坚实基础。请求结构化输出对于创建管道至关重要，其中语言模型的输出将作为后续系统或处理步骤的输入。

You can directly parse a JSON string from an LLM into a Pydantic object using the model_validate_json method. This is particularly useful as it combines parsing and validation in a single step.

| from pydantic import BaseModel, EmailStr, Field, ValidationError from typing import List, Optional from datetime import date # --- Pydantic Model Definition (from above) --- class User(BaseModel): name: str = Field(..., description="The full name of the user.") email: EmailStr = Field(..., description="The user's email address.") date_of_birth: Optional[date] = Field(None, description="The user's date of birth.") interests: List[str] = Field(default_factory=list, description="A list of the user's interests.") # --- Hypothetical LLM Output --- llm_output_json = """ { "name": "Alice Wonderland", "email": "alice.w@example.com", "date_of_birth": "1995-07-21", "interests": [ "Natural Language Processing", "Python Programming", "Gardening" ] } """ # --- Parsing and Validation --- try: # Use the model_validate_json class method to parse the JSON string. # This single step parses the JSON and validates the data against the User model. user_object = User.model_validate_json(llm_output_json) # Now you can work with a clean, type-safe Python object. print("Successfully created User object!") print(f"Name: {user_object.name}") print(f"Email: {user_object.email}") print(f"Date of Birth: {user_object.date_of_birth}") print(f"First Interest: {user_object.interests[0]}") # You can access the data like any other Python object attribute. # Pydantic has already converted the 'date_of_birth' string to a datetime.date object. print(f"Type of date_of_birth: {type(user_object.date_of_birth)}") except ValidationError as e: # If the JSON is malformed or the data doesn't match the model's types, # Pydantic will raise a ValidationError. print("Failed to validate JSON from LLM.") print(e) | | :—- |

您可以使用 model_validate_json 方法直接将来自 LLM 的 JSON 字符串解析为 Pydantic 对象。这特别高效，因为它在一个步骤中同时完成解析和验证。

This Python code demonstrates how to use the Pydantic library to define a data model and validate JSON data. It defines a User model with fields for name, email, date of birth, and interests, including type hints and descriptions. The code then parses a hypothetical JSON output from a Large Language Model (LLM) using the model_validate_json method of the User model. This method handles both JSON parsing and data validation according to the model’s structure and types. Finally, the code accesses the validated data from the resulting Python object and includes error handling for ValidationError in case the JSON is invalid.

这段 Python 代码演示了如何使用 Pydantic 库定义数据模型并验证 JSON 数据。它定义了一个包含姓名、电子邮件、出生日期和兴趣字段的 User 模型，附带类型提示和描述。代码随后使用 User 模型的 model_validate_json 方法解析来自大型语言模型（LLM）的假设 JSON 输出。该方法根据模型结构和类型要求处理 JSON 解析和数据验证。最后，代码从生成的 Python 对象中访问已验证数据，并包含 ValidationError 的异常处理，以应对 JSON 无效的情况。

For XML data, the xmltodict library can be used to convert the XML into a dictionary, which can then be passed to a Pydantic model for parsing. By using Field aliases in your Pydantic model, you can seamlessly map the often verbose or attribute-heavy structure of XML to your object’s fields.

对于 XML 数据，可使用 xmltodict 库将 XML 转换为字典，然后传递给 Pydantic 模型进行解析。通过在 Pydantic 模型中使用 Field 别名，可以无缝地将通常冗长或属性密集的 XML 结构映射到对象的字段。

This methodology is invaluable for ensuring the interoperability of LLM-based components with other parts of a larger system. When an LLM’s output is encapsulated within a Pydantic object, it can be reliably passed to other functions, APIs, or data processing pipelines with the assurance that the data conforms to the expected structure and types. This practice of “parse, don’t validate” at the boundaries of your system components leads to more robust and maintainable applications.

这种方法对于确保基于 LLM 的组件与更大系统其他部分的互操作性极为宝贵。当 LLM 输出封装在 Pydantic 对象中时，可以可靠地传递给其他函数、API 或数据处理管道，并确保数据符合预期结构和类型。这种在系统组件边界实施”解析而非验证”的原则，能够构建更健壮和可维护的应用程序。

Effectively utilizing system prompts, role assignments, contextual information, delimiters, and structured output significantly enhances the clarity, control, and utility of interactions with language models, providing a strong foundation for developing reliable agentic systems. Requesting structured output is crucial for creating pipelines where the language model’s output serves as the input for subsequent system or processing steps.

有效利用系统提示、角色分配、上下文信息、分隔符和结构化输出，显著提升了与语言模型交互的清晰度、控制力和实用性，为构建可靠的 Agentic 系统奠定了坚实基础。请求结构化输出对于创建管道至关重要，其中语言模型的输出将作为后续系统或处理步骤的输入。

Reasoning and Thought Process Techniques

推理与思维过程技术

Large language models excel at pattern recognition and text generation but often face challenges with tasks requiring complex, multi-step reasoning. This appendix focuses on techniques designed to enhance these reasoning capabilities by encouraging models to reveal their internal thought processes. Specifically, it addresses methods to improve logical deduction, mathematical computation, and planning.

大型语言模型擅长模式识别和文本生成，但在需要复杂多步骤推理的任务中常常面临挑战。本附录重点介绍旨在通过鼓励模型揭示其内部思维过程来增强推理能力的技术。具体而言，它探讨了改进逻辑推演、数学计算和规划的方法。

Chain of Thought (CoT)

思维链（Chain of Thought, CoT）

The Chain of Thought (CoT) prompting technique is a powerful method for improving the reasoning abilities of language models by explicitly prompting the model to generate intermediate reasoning steps before arriving at a final answer. Instead of just asking for the result, you instruct the model to “think step by step.” This process mirrors how a human might break down a problem into smaller, more manageable parts and work through them sequentially.

CoT helps the LLM generate more accurate answers, particularly for tasks that require some form of calculation or logical deduction, where models might otherwise struggle and produce incorrect results. By generating these intermediate steps, the model is more likely to stay on track and perform the necessary operations correctly.

There are two main variations of CoT:

Q: Sarah has 5 apples, and she buys 8 more. She eats 3 apples. How many apples does she have left? Let’s think step by step.
A: Let’s think step by step. Sarah starts with 5 apples. She buys 8 more, so she adds 8 to her initial amount: 5 + 8 = 13 apples. Then, she eats 3 apples, so we subtract 3 from the total: 13 - 3 = 10. Sarah has 10 apples left. The answer is 10.

CoT offers several advantages. It is relatively low-effort to implement and can be highly effective with off-the-shelf LLMs without requiring fine-tuning. A significant benefit is the increased interpretability of the model’s output; you can see the reasoning steps it followed, which helps in understanding why it arrived at a particular answer and in debugging if something went wrong. Additionally, CoT appears to improve the robustness of prompts across different versions of language models, meaning the performance is less likely to degrade when a model is updated. The main disadvantage is that generating the reasoning steps increases the length of the output, leading to higher token usage, which can increase costs and response time.

Best practices for CoT include ensuring the final answer is presented after the reasoning steps, as the generation of the reasoning influences the subsequent token predictions for the answer. Also, for tasks with a single correct answer (like mathematical problems), setting the model’s temperature to 0 (greedy decoding) is recommended when using CoT to ensure deterministic selection of the most probable next token at each step.

Self-Consistency

自洽性（Self-Consistency）

Building on the idea of Chain of Thought, the Self-Consistency technique aims to improve the reliability of reasoning by leveraging the probabilistic nature of language models. Instead of relying on a single greedy reasoning path (as in basic CoT), Self-Consistency generates multiple diverse reasoning paths for the same problem and then selects the most consistent answer among them.

Self-Consistency involves three main steps:

  1. Generating Diverse Reasoning Paths: The same prompt (often a CoT prompt) is sent to the LLM multiple times. By using a higher temperature setting, the model is encouraged to explore different reasoning approaches and generate varied step-by-step explanations.
  2. Extract the Answer: The final answer is extracted from each of the generated reasoning paths.
  3. Choose the Most Common Answer: A majority vote is performed on the extracted answers. The answer that appears most frequently across the diverse reasoning paths is selected as the final, most consistent answer.

This approach improves the accuracy and coherence of responses, particularly for tasks where multiple valid reasoning paths might exist or where the model might be prone to errors in a single attempt. The benefit is a pseudo-probability likelihood of the answer being correct, increasing overall accuracy. However, the significant cost is the need to run the model multiple times for the same query, leading to much higher computation and expense.

Step-Back Prompting

后退式提示（Step-Back Prompting）

Step-back prompting enhances reasoning by first asking the language model to consider a general principle or concept related to the task before addressing specific details. The response to this broader question is then used as context for solving the original problem.

This process allows the language model to activate relevant background knowledge and wider reasoning strategies. By focusing on underlying principles or higher-level abstractions, the model can generate more accurate and insightful answers, less influenced by superficial elements. Initially considering general factors can provide a stronger basis for generating specific creative outputs. Step-back prompting encourages critical thinking and the application of knowledge, potentially mitigating biases by emphasizing general principles.

Tree of Thoughts (ToT)

思维树（Tree of Thoughts, ToT）

Tree of Thoughts (ToT) is an advanced reasoning technique that extends the Chain of Thought method. It enables a language model to explore multiple reasoning paths concurrently, instead of following a single linear progression. This technique utilizes a tree structure, where each node represents a “thought”—a coherent language sequence acting as an intermediate step. From each node, the model can branch out, exploring alternative reasoning routes.

ToT is particularly suited for complex problems that require exploration, backtracking, or the evaluation of multiple possibilities before arriving at a solution. While more computationally demanding and intricate to implement than the linear Chain of Thought method, ToT can achieve superior results on tasks necessitating deliberate and exploratory problem-solving. It allows an agent to consider diverse perspectives and potentially recover from initial errors by investigating alternative branches within the “thought tree.”

These reasoning and thought process techniques are crucial for building agents capable of handling tasks that go beyond simple information retrieval or text generation. By prompting models to expose their reasoning, consider multiple perspectives, or step back to general principles, we can significantly enhance their ability to perform complex cognitive tasks within agentic systems.

Action and Interaction Techniques

行动与交互技术

Intelligent agents possess the capability to actively engage with their environment, beyond generating text. This includes utilizing tools, executing external functions, and participating in iterative cycles of observation, reasoning, and action. This section examines prompting techniques designed to enable these active behaviors.

Tool Use / Function Calling

工具使用/函数调用

A crucial ability for an agent is using external tools or calling functions to perform actions beyond its internal capabilities. These actions may include web searches, database access, sending emails, performing calculations, or interacting with external APIs. Effective prompting for tool use involves designing prompts that instruct the model on the appropriate timing and methodology for tool utilization.

Modern language models often undergo fine-tuning for “function calling” or “tool use.” This enables them to interpret descriptions of available tools, including their purpose and parameters. Upon receiving a user request, the model can determine the necessity of tool use, identify the appropriate tool, and format the required arguments for its invocation. The model does not execute the tool directly. Instead, it generates a structured output, typically in JSON format, specifying the tool and its parameters. An agentic system then processes this output, executes the tool, and provides the tool’s result back to the model, integrating it into the ongoing interaction.

User: What’s the weather like in London right now?

ReAct (Reason & Act)

ReAct（推理与行动）

ReAct, short for Reason and Act, is a prompting paradigm that combines Chain of Thought-style reasoning with the ability to perform actions using tools in an interleaved manner. ReAct mimics how humans operate – we reason verbally and take actions to gather more information or make progress towards a goal.

The ReAct pattern involves a loop:

  1. Thought: The model generates a thought process, explaining its current understanding and plan.
  2. Action: Based on the thought, the model decides to perform an action, often using a tool (e.g., Search, Calculator, API call). It outputs the name of the tool and the required input.
  3. Observation: The agentic system executes the tool and provides the result (Observation) back to the model.
  4. The loop continues with the model generating a new “Thought” based on the latest “Observation,” leading to further “Actions” and “Observations” until the task is complete and the model outputs a “Final Answer.”

This interleaved process of thinking and acting allows the agent to dynamically gather information, react to tool outputs, and refine its approach, making it particularly effective for tasks requiring interaction with dynamic environments or external knowledge sources.

Thought: The user is asking for two pieces of information: the capital of France and its population. I can use a search tool to find this information. I will start by confirming the capital, although I am likely to know it.
Action: Search({“query”: “capital of France”})
Observation: Paris

Thought: Okay, the capital is Paris. Now I need to find the population of Paris.
Action: Search({“query”: “current population of Paris”})
Observation: The population of Paris was estimated to be around 2.1 million people in 2023.

Thought: I have found both pieces of information. I can now provide the final answer to the user.
Final Answer: The capital of France is Paris, and its estimated population in 2023 was around 2.1 million people.

These techniques are vital for building agents that can actively engage with the world, retrieve real-time information, and perform tasks that require interacting with external systems.

Advanced Techniques

高级技术

Beyond the foundational, structural, and reasoning patterns, there are several other prompting techniques that can further enhance the capabilities and efficiency of agentic systems. These range from using AI to optimize prompts to incorporating external knowledge and tailoring responses based on user characteristics.

Automatic Prompt Engineering (APE)

自动提示工程（APE）

Recognizing that crafting effective prompts can be a complex and iterative process, Automatic Prompt Engineering (APE) explores using language models themselves to generate, evaluate, and refine prompts. This method aims to automate the prompt writing process, potentially enhancing model performance without requiring extensive human effort in prompt design.

The general idea is to have a “meta-model” or a process that takes a task description and generates multiple candidate prompts. These prompts are then evaluated based on the quality of the output they produce on a given set of inputs (perhaps using metrics like BLEU or ROUGE, or human evaluation). The best-performing prompts can be selected, potentially refined further, and used for the target task. Using an LLM to generate variations of a user query for training a chatbot is an example of this.

Of course. Here is a rephrased and slightly expanded explanation of programmatic prompt optimization using frameworks like DSPy:

Another powerful prompt optimization technique, notably promoted by the DSPy framework, involves treating prompts not as static text but as programmatic modules that can be automatically optimized. This approach moves beyond manual trial-and-error and into a more systematic, data-driven methodology.

The core of this technique relies on two key components:

  1. A Goldset (or High-Quality Dataset): This is a representative set of high-quality input-and-output pairs. It serves as the “ground truth” that defines what a successful response looks like for a given task.
  2. An Objective Function (or Scoring Metric): This is a function that automatically evaluates the LLM’s output against the corresponding “golden” output from the dataset. It returns a score indicating the quality, accuracy, or correctness of the response.

Using these components, an optimizer, such as a Bayesian optimizer, systematically refines the prompt. This process typically involves two main strategies, which can be used independently or in concert:

The ultimate goal for both strategies is to maximize the scores from the objective function, effectively “training” the prompt to produce results that are consistently closer to the high-quality goldset. By combining these two approaches, the system can simultaneously optimize what instructions to give the model and which examples to show it, leading to a highly effective and robust prompt that is machine-optimized for the specific task.

Iterative Prompting / Refinement

迭代式提示/优化

This technique involves starting with a simple, basic prompt and then iteratively refining it based on the model’s initial responses. If the model’s output isn’t quite right, you analyze the shortcomings and modify the prompt to address them. This is less about an automated process (like APE) and more about a human-driven iterative design loop.

Providing Negative Examples

提供反例

While the principle of “Instructions over Constraints” generally holds true, there are situations where providing negative examples can be helpful, albeit used carefully. A negative example shows the model an input and an undesired output, or an input and an output that should not be generated. This can help clarify boundaries or prevent specific types of incorrect responses.

Example of what NOT to do:
Input: List popular landmarks in Paris.
Output: The Eiffel Tower, The Louvre, Notre Dame Cathedral.

Using Analogies

使用类比

Framing a task using an analogy can sometimes help the model understand the desired output or process by relating it to something familiar. This can be particularly useful for creative tasks or explaining complex roles.

Factored Cognition / Decomposition

分解式认知/任务拆解

For very complex tasks, it can be effective to break down the overall goal into smaller, more manageable sub-tasks and prompt the model separately on each sub-task. The results from the sub-tasks are then combined to achieve the final outcome. This is related to prompt chaining and planning but emphasizes the deliberate decomposition of the problem.

Retrieval Augmented Generation (RAG)

检索增强生成（RAG）

RAG is a powerful technique that enhances language models by giving them access to external, up-to-date, or domain-specific information during the prompting process. When a user asks a question, the system first retrieves relevant documents or data from a knowledge base (e.g., a database, a set of documents, the web). This retrieved information is then included in the prompt as context, allowing the language model to generate a response grounded in that external knowledge. This mitigates issues like hallucination and provides access to information the model wasn’t trained on or that is very recent. This is a key pattern for agentic systems that need to work with dynamic or proprietary information.

Persona Pattern (User Persona):

角色模式（用户画像）

While role prompting assigns a persona to the model, the Persona Pattern involves describing the user or the target audience for the model’s output. This helps the model tailor its response in terms of language, complexity, tone, and the kind of information it provides.

Explain quantum physics: [Insert basic explanation request]

These advanced and supplementary techniques provide further tools for prompt engineers to optimize model behavior, integrate external information, and tailor interactions for specific users and tasks within agentic workflows.

Using Google Gems

使用 Google Gems

Google’s AI “Gems” (see Fig. 1) represent a user-configurable feature within its large language model architecture. Each “Gem” functions as a specialized instance of the core Gemini AI, tailored for specific, repeatable tasks. Users create a Gem by providing it with a set of explicit instructions, which establishes its operational parameters. This initial instruction set defines the Gem’s designated purpose, response style, and knowledge domain. The underlying model is designed to consistently adhere to these pre-defined directives throughout a conversation.

This allows for the creation of highly specialized AI agents for focused applications. For example, a Gem can be configured to function as a code interpreter that only references specific programming libraries. Another could be instructed to analyze data sets, generating summaries without speculative commentary. A different Gem might serve as a translator adhering to a particular formal style guide. This process creates a persistent, task-specific context for the artificial intelligence.

Consequently, the user avoids the need to re-establish the same contextual information with each new query. This methodology reduces conversational redundancy and improves the efficiency of task execution. The resulting interactions are more focused, yielding outputs that are consistently aligned with the user’s initial requirements. This framework allows for applying fine-grained, persistent user direction to a generalist AI model. Ultimately, Gems enable a shift from general-purpose interaction to specialized, pre-defined AI functionalities.


Fig.1: Example of Google Gem usage.

Using LLMs to Refine Prompts (The Meta Approach)

We’ve explored numerous techniques for crafting effective prompts, emphasizing clarity, structure, and providing context or examples. This process, however, can be iterative and sometimes challenging. What if we could leverage the very power of large language models, like Gemini, to help us improve our prompts? This is the essence of using LLMs for prompt refinement – a “meta” application where AI assists in optimizing the instructions given to AI.

This capability is particularly “cool” because it represents a form of AI self-improvement or at least AI-assisted human improvement in interacting with AI. Instead of solely relying on human intuition and trial-and-error, we can tap into the LLM’s understanding of language, patterns, and even common prompting pitfalls to get suggestions for making our prompts better. It turns the LLM into a collaborative partner in the prompt engineering process.

How does this work in practice? You can provide a language model with an existing prompt that you’re trying to improve, along with the task you want it to accomplish and perhaps even examples of the output you’re currently getting (and why it’s not meeting your expectations). You then prompt the LLM to analyze the prompt and suggest improvements.

A model like Gemini, with its strong reasoning and language generation capabilities, can analyze your existing prompt for potential areas of ambiguity, lack of specificity, or inefficient phrasing. It can suggest incorporating techniques we’ve discussed, such as adding delimiters, clarifying the desired output format, suggesting a more effective persona, or recommending the inclusion of few-shot examples.

The benefits of this meta-prompting approach include:

It’s important to note that the LLM’s suggestions are not always perfect and should be evaluated and tested, just like any manually engineered prompt. However, it provides a powerful starting point and can significantly streamline the refinement process.

Existing Prompt:
“Summarize the main points and list important names and places from this article: [insert article text]”

Suggestions for Improvement:

In this example, we’re using the LLM to critique and enhance another prompt. This meta-level interaction demonstrates the flexibility and power of these models, allowing us to build more effective agentic systems by first optimizing the fundamental instructions they receive. It’s a fascinating loop where AI helps us talk better to AI.

Prompting for Specific Tasks

While the techniques discussed so far are broadly applicable, some tasks benefit from specific prompting considerations. These are particularly relevant in the realm of code and multimodal inputs.

Code Prompting

Language models, especially those trained on large code datasets, can be powerful assistants for developers. Prompting for code involves using LLMs to generate, explain, translate, or debug code. Various use cases exist:

Effective code prompting often requires providing sufficient context, specifying the desired language and version, and being clear about the functionality or issue.

Multimodal Prompting

While the focus of this appendix and much of current LLM interaction is text-based, the field is rapidly moving towards multimodal models that can process and generate information across different modalities (text, images, audio, video, etc.). Multimodal prompting involves using a combination of inputs to guide the model. This refers to using multiple input formats instead of just text.

As multimodal capabilities become more sophisticated, prompting techniques will evolve to effectively leverage these combined inputs and outputs.

Best Practices and Experimentation

Becoming a skilled prompt engineer is an iterative process that involves continuous learning and experimentation. Several valuable best practices are worth reiterating and emphasizing:

Prompt engineering is a skill that improves with practice. By applying these principles and techniques, and by maintaining a systematic approach to experimentation and documentation, you can significantly enhance your ability to build effective agentic systems.

Conclusion

结论

This appendix provides a comprehensive overview of prompting, reframing it as a disciplined engineering practice rather than a simple act of asking questions. Its central purpose is to demonstrate how to transform general-purpose language models into specialized, reliable, and highly capable tools for specific tasks. The journey begins with non-negotiable core principles like clarity, conciseness, and iterative experimentation, which are the bedrock of effective communication with AI. These principles are critical because they reduce the inherent ambiguity in natural language, helping to steer the model’s probabilistic outputs toward a single, correct intention. Building on this foundation, basic techniques such as zero-shot, one-shot, and few-shot prompting serve as the primary methods for demonstrating expected behavior through examples. These methods provide varying levels of contextual guidance, powerfully shaping the model’s response style, tone, and format. Beyond just examples, structuring prompts with explicit roles, system-level instructions, and clear delimiters provides an essential architectural layer for fine-grained control over the model.

本附录全面概述了提示工程，将其重新定位为一项有纪律的工程实践，而非简单的提问行为。其核心目标是展示如何将通用语言模型转化为针对特定任务的专业化、可靠且高度能干的工具。这一旅程始于不容妥协的核心原则：清晰性、简洁性和迭代实验，这些是与 AI 有效沟通的基石。这些原则至关重要，因其减少了自然语言固有的歧义，帮助引导模型的概率输出朝向明确正确的意图。在此基础上，基础技术（如 zero-shot、one-shot 和 few-shot 提示）作为通过示例展示预期行为的主要手段。这些方法提供不同层级的上下文指导，有力塑造模型的响应风格、语气和格式。超越示例范畴，使用明确角色、系统级指令和清晰分隔符构建提示，为精细控制模型提供了必要的架构层次。

The importance of these techniques becomes paramount in the context of building autonomous agents, where they provide the control and reliability necessary for complex, multi-step operations. For an agent to effectively create and execute a plan, it must leverage advanced reasoning patterns like Chain of Thought and Tree of Thoughts. These sophisticated methods compel the model to externalize its logical steps, systematically breaking down complex goals into a sequence of manageable sub-tasks. The operational reliability of the entire agentic system hinges on the predictability of each component’s output. This is precisely why requesting structured data like JSON, and programmatically validating it with tools such as Pydantic, is not a mere convenience but an absolute necessity for robust automation. Without this discipline, the agent’s internal cognitive components cannot communicate reliably, leading to catastrophic failures within an automated workflow. Ultimately, these structuring and reasoning techniques are what successfully convert a model’s probabilistic text generation into a deterministic and trustworthy cognitive engine for an agent.

这些技术在构建自主智能体的背景下变得至关重要，它们为复杂、多步骤操作提供了必要的控制与可靠性。为使智能体能够有效地制定和执行计划，必须利用高级推理模式，如思维链和思维树。这些复杂方法迫使模型外化其逻辑步骤，系统地将复杂目标分解为可管理的子任务序列。整个智能体系统的运营可靠性依赖于各组件输出的可预测性。这正是为何请求 JSON 等结构化数据，并使用 Pydantic 等工具进行程序化验证，不仅是一种便利，更是强大自动化的绝对必要条件。缺乏这种纪律，智能体的内部认知组件无法可靠通信，导致自动化工作流中的灾难性故障。最终，这些结构化与推理技术成功将模型的概率文本生成转化为智能体可靠且确定的认知引擎。

Furthermore, these prompts are what grant an agent its crucial ability to perceive and act upon its environment, bridging the gap between digital thought and real-world interaction. Action-oriented frameworks like ReAct and native function calling are the vital mechanisms that serve as the agent’s hands, allowing it to use tools, query APIs, and manipulate data. In parallel, techniques like Retrieval Augmented Generation (RAG) and the broader discipline of Context Engineering function as the agent’s senses. They actively retrieve relevant, real-time information from external knowledge bases, ensuring the agent’s decisions are grounded in current, factual reality. This critical capability prevents the agent from operating in a vacuum, where it would be limited to its static and potentially outdated training data. Mastering this full spectrum of prompting is therefore the definitive skill that elevates a generalist language model from a simple text generator into a truly sophisticated agent, capable of performing complex tasks with autonomy, awareness, and intelligence.

此外，这些提示赋予智能体感知环境并与之交互的关键能力，弥合了数字思维与现实世界行动之间的鸿沟。ReAct 和原生函数调用等面向行动的框架是智能体双手的重要机制，使其能够使用工具、查询 API 和操作数据。同时，检索增强生成（RAG）及更广泛的上下文工程学科充当智能体的感官。它们主动从外部知识库检索相关的实时信息，确保智能体的决策基于当前的事实现实。这种关键能力防止智能体在真空中运作，使其不局限于静态且可能过时的训练数据。因此，掌握完整的提示技术谱系，是将通用语言模型从简单文本生成器提升为真正复杂智能体的决定性技能，使其能够自主、情境感知地智能执行复杂任务。

References

参考文献

Here is a list of resources for further reading and deeper exploration of prompt engineering techniques:

  1. Prompt Engineering, https://www.kaggle.com/whitepaper-prompt-engineering
  2. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models, https://arxiv.org/abs/2201.11903
  3. Self-Consistency Improves Chain of Thought Reasoning in Language Models, https://arxiv.org/pdf/2203.11171
  4. ReAct: Synergizing Reasoning and Acting in Language Models, https://arxiv.org/abs/2210.03629
  5. Tree of Thoughts: Deliberate Problem Solving with Large Language Models, https://arxiv.org/pdf/2305.10601
  6. Take a Step Back: Evoking Reasoning via Abstraction in Large Language Models, https://arxiv.org/abs/2310.06117
  7. DSPy: Programming—not prompting—Foundation Models https://github.com/stanfordnlp/dspy