Prompt Engineering for Large Models: The Science and Art of Building Large Model Applications

From Prompt Design to Evaluable Large-Scale Model Applications

做过知识库问答、文本整理或smart (phone, system, bomb etc)助手的人，往往会遇到同一个问题：一条提示词在演示中表现不错，换一批问题、加几段材料，结果就不稳定了。《大模型提示工程：构建大模型应用的科学与艺术》适合带着这类问题阅读。它把提示工程放进完整的应用流程，讨论输入怎样组织、上下文怎样取舍、模型输出怎样接入后续步骤，以及改动后如何判断效果。

The original title of this book is Prompt Engineering for Large Language Models, authored by John Berryman and Albert Ziegler, both of whom were involved in the early development of GitHub Copilot. The Simplified Chinese edition was translated by Ye Zhiyuan and Wang Gang and published by China Machine Press in June 2026, with ISBN 9787111810117. Edition details and the complete table of contents can be viewed via the publisher’s book page link at the top of this page.

Content Structure: Expanding outward from a single model call

First, understand the model; then define the tasks in the application.

The foundational section begins with morphemes, text generation, and dialogue models, before moving on to application design and quality evaluation. As you read, it’s worth repeatedly asking yourself: How can user needs be transformed into input that the model can process? How can the generated text be turned into results that users can actually use? Thinking through these two conversion steps is usually more helpful than repeatedly revising a single command.

Treat the context as input for the design

The technical section focuses on prompt content, fragment assembly, output control, and model selection, covering few-shot examples, RAG, summarization, and the spatial relationships and dependencies among pieces of information. The most valuable perspective to bring back to the project here is to examine the purpose of each piece of material: Is it meant to explain the task, provide evidence, or specify the output format? When time is limited or sources conflict, this breakdown can help the team discuss what to omit and what to retain.

From Single Responses to Agents and Evaluation

The advanced section continues the discussion on tool usage, conversational context, task workflows, and offline and online evaluation. Readers can compare these concepts with their own processes to determine which steps should be left to the model to decide and which are best handled using explicit rules, and then examine how errors might affect the next step. For the chapter structure and topics, seeO’Reilly Original Catalog.

Who is this book for, and what background knowledge is required?

• Developers who are already able to call the model API:It is suitable for organizing prompt templates, retrieving materials, structuring the relationships between outputs and test sets, and gradually improving existing prototypes.
• Technical Product Managers Involved in AI Product Design:You may want to focus on the sections on task design, workflows, and evaluation to help ensure that requirements discussions result in clearly defined inputs, outputs, and acceptance criteria.
• People responsible for the knowledge base or automated processes:It’s best to read this while referring to real-world examples to determine whether failures stem from insufficient data, ambiguous task boundaries, or incomplete output processing.

Before reading, it is best to have a basic understanding of APIs, JSON, and basic programming workflows; the practical sections require the ability to save input and output, modify call parameters, and compare results. While you can read the conceptual chapters even without development experience, applying the methods in practice still requires a foundation in programming and API calls. The original official version is intended for intermediate to advanced readers; when selecting this book, you should assess your own project experience.

How to Use It in Conjunction with *AI Engineering* and *Illustrated Guide to Large Models*

Below, we compare the three books based on learning objectives to help you decide which one to read first. While there is some overlap in their content, the main differences lie in their focus.

If you’re trying to figure out why your model’s performance fluctuates, start by reading this book; if your model is already running smoothly and your next steps involve deployment, evaluation, and cost control, move on to *AI Engineering*; if you still lack a comprehensive understanding of concepts such as vector representations, attention mechanisms, and retrieval, *Illustrated Large Models* is a good reference to consult alongside this book.

What is the recommended way to read so that you can develop reusable methods?

1. Choose a task that's small enough.For example, when organizing product information into fixed fields, first define the input ranges, required fields, and how to handle cases where the value cannot be determined.
2. Keep a set of fixed examples.In addition to standard issues, consider cases involving missing information, conflicting data, and scope deviations; throughout the review process, consistently use the same set of examples to compare proposals.
3. Change only one type of factor at a time.Try adjusting the examples, the order of the context, or the output constraints one at a time, while recording quality, response time, and call overhead, to avoid drawing conclusions based on a single response.
4. Turn failure cases into check items.When workflows become lengthy, systematically review the retrieval, generation, and post-processing stages so that corrections made once can be reused later.

Recommendation Value and Reading Boundaries

This book is designed to help readers develop a more structured approach to their work: when encountering performance issues, first break down the input, context, model behavior, and evaluation methods before deciding where to make changes. For readers who need to iterate on AI applications repeatedly, this analytical framework is more easily transferable to new tasks than simply collecting a large number of ready-made prompt templates.

When studying, be sure to refer to the official documentation for the current model. Tool calls, structured outputs, and model inference interfaces are subject to change, so the strategies described in this book must be revalidated for your specific tasks. Prompt engineering alone cannot guarantee factual accuracy, data security, or the reliability of business processes; for applications involving external operations, you should also implement access controls, result validation, and necessary manual verification.