Everything you need to know about 'prompt engineering' and 'context engineering' from a prompt engineer who has led various LLM projects in the field.
Starting from the fundamental question, "What is a prompt?", this book closely examines the role of a prompt engineer in the design of model, data, and service pipelines. It also analyzes the duties and capabilities of a prompt engineer, not only in the area of ​​prompting techniques but also in the real world, including client business analysis, planning, and experimentation.
This book will serve as a guidebook that covers both theory and practice for those who want to establish work processes as a prompt engineer practitioner, PMs or planners who work with prompt engineers, HR managers who hire prompt engineers, and aspiring prompt engineers.

The quality and accuracy of LLM answers vary greatly depending on the prompts provided.
Therefore, in order to derive the most ideal LLM answers, we need to design, experiment with, and refine prompts. This is called 'prompt engineering.'
And 'prompt engineer' is a job that specializes in this type of prompt engineering.
In other words, a prompt engineer is someone who finds ways for LLM to reliably provide the intended answers.
However, ending the explanation like this leaves a lot of room for misunderstanding.
Typically, the job of a prompt engineer is known to be to design prompts that maximize the LLM's answers.
However, in practice, prompt engineers are responsible for a wide range of tasks, including defining the information required for the model and designing a prompt design accordingly, analyzing the client's business, designing service pipelines, designing data, and verifying model performance.

--- pp.25-26

I have worked on a project that provides a service to answer questions related to labor law.
Legal counseling is a process of understanding the counselor's situation and reaching a conclusion on how to resolve the case based on the logical system established by the law.
Therefore, LLM's response in the project was designed to ① understand the user's situation, ② explain the content of the relevant legal provisions, ③ explain how the law can be applied to the user's situation, and ④ summarize the conclusion.
However, the client pointed out that the format of the answer generation, where the conclusion is placed at the end of the answer, is inconvenient because the user has to read a lot of text first to understand the conclusion, and requested that it be changed to a format of the answer generation, where the conclusion is created as a summary first and the explanation is placed afterward.
… Legal consultation is a task that requires logical thinking to reach a conclusion through the following stages: ① ascertaining the facts, ② selecting relevant legal provisions, ③ examining whether the provisions apply, and ④ drawing a conclusion.
However, asking the model to immediately infer the ④ conclusion for such content requires a higher level of difficult inference than inference step by step, so the final correct answer rate is bound to be lower.
Therefore, if you create a summary first for a two-part answer, it is more likely to be wrong because it requires advanced reasoning.
…
For this reason, the CoT (Chain of Thought) technique, a prompting technique that divides into logical steps and instructs the model to generate content that is relatively easy to infer step by step to reach a conclusion, is further increasing the model's accuracy rate. The Deepseek R1 model, which has recently become an issue, is also increasing the accuracy rate in questions about inference in this way.

--- pp.132-134

The slogan, “The core competency of AI is now context engineering, not prompting,” fundamentally presupposes the use of powerful models, but limitations still exist at the current level, where it is difficult to say that any model has achieved AGI.
Furthermore, in the actual market, there is a strong demand for models with low parameters to keep maintenance costs low, and this demand is expected to increase as models become more sophisticated.
In this situation, it does not seem possible to immediately free ourselves from prompt engineering.
So I disagree with the opinion that prompt engineering is no longer important.
However, I don't think the importance of context engineering capabilities is overrated.
Ultimately, I believe that prompt engineers in AI agent projects need both the ability to see the forest (context engineering ability) and the ability to grow the trees (prompt engineering ability).

--- pp.192-193