As artificial intelligence develops rapidly, prompts, as a key tool for interacting with large language models (LLMs), are gradually attracting more and more attention. Prompt engineering is not only the exclusive domain of technical developers, but ordinary users can also communicate better with intelligent systems through effective prompts. This article will introduce the concept and importance of prompt engineering, as well as how to use zero-sample prompts, few-sample prompts, and thought chains to improve the response effect of the model.

2. What is the prompt word project?

Prompt word engineering is a discipline that studies how to design and optimize prompt words. Simply put, prompt words are the instructions you give to the big model. By properly designing prompt words, the quality and accuracy of the model output can be significantly improved. With the widespread application of big language models, the use of prompt words is not limited to simple questions and answers. Developers also seek to make the model perform better in complex application scenarios.

In the prompting project, researchers found that different prompts directly affect the response effect of the model. In order for the model to better understand and perform tasks, the structure, content and style of the prompts must be carefully designed. This technology originally came from the field of natural language processing (NLP), and a large number of research papers have explored effective prompt writing methods. For easy learning, readers can visit the Prompting Guide website to learn more about related knowledge and techniques.

3. Classification of prompt words

In the prompt word engineering, prompt words are mainly divided into two categories: zero-shot prompts and few-shot prompts. Thought chain is an advanced usage of few-shot prompts.

Note: The reason why the big model can directly help us solve various problems is that its theoretical basis is zero-shot prompts. After training, the big model has the ability to understand user instructions and respond appropriately. When the results given by the big model do not meet the user's expectations, we should consider adjusting the feedback of the big model through small sample prompts.

3.1 Zero-Shot Prompting

Zero-shot prompting means that the model generates a response based only on the task description without any examples. In this case, the user only needs to provide a clear description to the model, and the model will understand the task and generate the corresponding results. As the training and instruction adjustment capabilities of large language models continue to improve, the effect of zero-shot prompting is becoming more and more significant.

However, zero-shot hints may not work well on some complex tasks. If the model fails to understand the task correctly or the generated content deviates from expectations, it is recommended to provide some examples, which introduces the concept of few-shot hints.

3.2 Few-Shot Prompting

Few-shot prompts provide examples in the prompt to guide the model to generate better responses. While large language models perform well in zero-shot settings, few-shot prompts can significantly improve their performance when faced with complex tasks. By providing relevant examples, users can help the model better understand the task requirements, thereby improving the accuracy and coherence of the output.

In some cases, if zero-shot and few-shot prompts still do not meet your needs, you may need to consider more complex prompting strategies, such as the Chain of Thought technique.

Case 1: Text Classification

This is the Few Shot example that was often mentioned in the past two years. The essence of Few Shot is to provide a small number of learning samples to the big model so that the big model can understand the user's expectations or tendencies through sample learning before responding to the user's questions, so as to give the user a more satisfactory answer.

Task Description : Classify the following sentences as "positive", "negative" or "neutral".

sentence :

1. “This movie is amazing!”
2. “I’m disappointed with this trip.”
3. “The weather is nice.”

Less sample tips :

Please categorize the sentences according to the following examples:

1. “This movie is amazing!” → Positive

2. “I was disappointed with this trip.” → Negative

3. "The weather is nice." → Neutral

Now, please classify the following sentences:

1. “I am very happy with my work today.”

Model Output : By providing the previous example, the model can better understand the sentiment of the sentence and outputs “Positive”.

Case Study

In this case, by giving a small number of examples, the model was able to learn how to classify the sentiment of the sentence. Compared to zero-shot prompts, few-shot prompts perform better on complex tasks such as sentiment analysis because the model can learn from the context of the examples and make more accurate judgments.

Case 2: Recommended recipes

Task description : Recommend suitable simple recipes based on the ingredients provided by the user.

Recommended ingredients :

• egg
• tomato
• bread

Less sample tips :

Please recommend a suitable recipe based on the following examples:

1. Ingredients: eggs, milk, flour

   Recommended recipe: Omelette - Mix eggs, milk and flour, fry into pancakes and top with your favorite seasonings.

2. Ingredients: tomatoes, pasta, olive oil

   Recommended recipe: Tomato Pasta - Dice tomatoes, cook in olive oil, add to cooked pasta, season and enjoy.

Now, please recommend a recipe based on the following ingredients:

Model output:

1. Ingredients: eggs, tomatoes, bread Recommended recipe: Tomato and egg sandwich - fry the eggs and sandwich them with sliced ​​tomatoes in bread, simple and delicious.
   

Case Study

In this Few-Shot case, we provide two specific ingredients and corresponding simple recipes to help the model learn how to make recommendations based on the ingredients provided by the user. Each example clearly shows how to convert the ingredients into specific recipes, allowing users to understand how the model recommends.

3.3 Thinking Chain: Technology to Improve Reasoning Ability

Chain of Thought (CoT) is a prompt strategy that breaks down complex questions into multiple intermediate steps, aiming to enhance the reasoning ability of language models. By providing a series of intermediate reasoning processes in the prompt, users can guide the model to generate more logical and accurate answers.

Practical tips: Add one more sentence to the prompt: "Please think step by step" to use the thinking chain.

The following is a paragraph that may not be easy to understand, so I marked it in gray. In actual use, just remember one thing: if the problem you ask the big model to help you solve involves a certain logical reasoning process, and the big model's response to you is not very accurate, just add one more sentence in the prompt: "Please think step by step" , and you will find that it becomes smarter instantly.

How Thought Chaining Works

1. Decompose the problem : Break down a complex problem into smaller sub-problems. For example, in the math problem "chickens and rabbits are in the same cage, there are 35 heads and 94 legs, how many chickens and rabbits are there?", you can first assume that there are only chickens, calculate the number of legs, and then compare it with the actual number of legs to get the difference in the number of legs, and gradually deduce the number of chickens and rabbits.

2. Guided reasoning : By presenting these intermediate steps in order, the model is guided to reason and calculate along the set thinking path. Based on these prompts, the model will combine its own knowledge and language understanding ability to process each step in turn and generate the corresponding intermediate results and final answers.

Application scenarios of thinking chain

• Mathematical reasoning : When solving mathematical problems, thought chains can help the model gradually demonstrate the solution ideas and improve the accuracy and explainability of calculations. For example, in geometric proofs and algebraic equation solutions, conclusions can be drawn through step-by-step reasoning.

• Logical reasoning : For logical puzzles and reasoning problems, the thinking chain can guide the model to conduct logical analysis, making the reasoning process clearer and more reasonable.

• Multi-step tasks : When dealing with tasks that require multi-step operations, such as making a plan or solving a complex engineering problem, thought chaining can help the model complete the task step by step in a reasonable order.

Case 1: Chickens and rabbits in the same cage

Problem description: There are chickens and rabbits in the same cage. There are 35 heads and 94 legs. How many chickens and rabbits are there? Please think step by step to solve this problem .

When ChatGPT was first launched, we found that the big model was a liberal arts student with poor math skills. It would even get elementary school math problems like chickens and rabbits in the same cage wrong. However, as long as we used the thinking chain routine and let it think step by step, it would give the correct answer.

The following is the output of the above problem. Since we let the big model think step by step, it will output as follows:

Step 1: Assuming all animals are chickens, the total number of legs of chickens is 35 × 2 = 70. Step 2: Now we know that there are actually 94 legs, so the difference in the number of legs is 94 - 70 = 24 legs. Step 3: Each rabbit has 2 more legs than a chicken, so the number of rabbits is 24 ÷ (4 - 2) = 12 rabbits. Step 4: Finally, the number of chickens is the total number of chickens minus the number of rabbits, which is 35 - 12 = 23 chickens.

From this we can conclude that there are 23 chickens and 12 rabbits.

Case Study

In this case, the chain of thought helps the model break down complex problems into multiple simple steps. Through step-by-step reasoning, the model can not only give the final answer, but also show the problem-solving process, making the results more explainable.

Case 2

Task description : A team is planning a project. Team members need to select appropriate personnel to form a project team based on their respective skills, available time and project requirements. Please infer the final project team members.

Known Information

1. Team members, their skills and availability :

• Xiao Li: Skills (programming, design), available time (20 hours per week)
• Xiao Zhang: Skills (programming, data analysis), available time (15 hours per week)
• Xiao Wang: Skills (design, marketing), available time (25 hours per week)
• Xiao Zhao: Skills (data analysis, project management), available time (10 hours per week)

• Project requirements :

• Programming: At least 2 people
• Design: At least 1 person
• Data analysis: at least 1 person
• Project management: at least 1 person
• Required skills:
• Available time: Total project working time is 50 hours per week

The large model will output the following reasoning process

1. First, I need to determine the skill requirements and available time for the project.

• Project requirements: Programming (at least 2 people), Design (at least 1 person), Data Analysis (at least 1 person), Project Management (at least 1 person)
• Total available time: 50 hours

• Next, I will analyze each member’s skills and available time:

• Xiao Li: Programming, design, 20 hours
• Xiao Zhang: Programming, data analysis, 15 hours
• Xiao Wang: Design, marketing, 25 hours
• Xiao Zhao: Data analysis, project management, 10 hours

• According to the project requirements, I need to find qualified members:

• Programming: Both Xiaoli and Xiaozhang have the skills, so choosing them can meet the requirement (20 + 15 = 35 hours).
• Design: Xiao Li and Xiao Wang both have the skills, and choosing Xiao Wang can meet the requirements (25 hours).
• Data analysis: Xiao Zhang or Xiao Zhao both have the skills, so choosing Xiao Zhang can meet the requirement (15 hours).
• Project Management: Xiao Zhao has the skills, so choosing Xiao Zhao can meet the requirement (10 hours).

• Now I calculate the total available time of the selected members:

• Xiao Li: 20 hours
• Xiao Zhang: 15 hours
• Xiao Wang: 25 hours
• Xiao Zhao: 10 hours
• Total time = 20 + 15 + 25 + 10 = 70 hours

• I found that the total available time (70 hours) exceeded the 50 hours required for the project, so I needed to adjust the team members.

• If Xiao Li (20 hours) and Xiao Zhang (15 hours) are selected for programming, Xiao Wang (25 hours) can be replaced by Xiao Zhao (10 hours), so the total time will be reduced:
• New combination: Xiao Li (20 hours), Xiao Zhang (15 hours), Xiao Zhao (10 hours). Total time = 20 + 15 + 10 = 45 hours.

• The final project team members are:

• Xiao Li (Programming)
• Xiao Zhang (Programming, Data Analysis)
• Xiao Zhao (Data Analysis, Project Management)

• verify:

• Programming: Xiao Li and Xiao Zhang (meet the requirements)
• Design: Xiao Wang was not selected, and other designers need to be found (does not meet the requirements, but external resources can be considered)
• Data analysis: Xiao Zhang and Xiao Zhao (meet the requirements)
• Project Management: Xiao Zhao (meets the requirements)
• Total time: 45 hours (as required)

Large model output results

Through the above chain of thought reasoning process, we finally draw the conclusion:

• Final project team members :

• Xiao Li (Programming)
• Xiao Zhang (Programming, Data Analysis)
• Xiao Zhao (Data Analysis, Project Management)

• Note : Design skills are not fully met, external resources may be considered.

Case Study

In this chain of thought example, AI needs to go through multiple steps of reasoning to determine the members of the project team. By analyzing each member's skills, available time, and project requirements, AI is able to gradually narrow down the options and ultimately determine a team combination that meets most of the requirements.

This chain of thought structure helps users sort out complex logical relationships and ensure that each step is based on reasonable reasoning. It shows how to make appropriate decisions through step-by-step analysis when faced with multiple conditions and constraints.

Hopefully this complex thought chain example can help you better understand the application of thought chains! If you have other needs or want further examples, please feel free to let me know.

4. Practical application of prompt word engineering

The application scope of the prompt word project is wide, covering education, medical care, finance, customer service and other fields. In education, using the prompt word project, teachers can provide students with clearer learning guidance and help them reason step by step when solving problems. In the medical field, prompt words can help doctors obtain patient information and medical records more effectively and improve diagnosis efficiency.

4.1 Education

In the field of education, the prompt word project can help teachers design more effective teaching plans. By providing clear task descriptions and examples, teachers can guide students to solve problems step by step and enhance their logical thinking ability. For example, teachers can use the thinking chain strategy to guide students to analyze math problems step by step and help them master problem-solving skills.

4.2 Medical field

In the medical field, doctors can use prompt word engineering to quickly obtain patients' medical records. With carefully designed prompt words, doctors can interact with intelligent systems more effectively, thereby improving diagnostic efficiency and accuracy. For example, when asking about a patient's condition, doctors can use a few sample prompts and provide some typical medical records as references to help the system better understand the patient's symptoms.

4.3 Customer Service Area

In the customer service field, prompt word engineering can help customer service personnel quickly solve customer problems. By designing effective prompt words, customer service personnel can guide the intelligent customer service system to understand customer needs more accurately and provide better services. For example, customer service personnel can use the thought chain strategy to guide the system to gradually analyze customer problems and provide more accurate solutions.

5. Tips: How to write good prompt words

Writing good prompts is not difficult, but it requires some skills and experience. Here are some practical tips to help you achieve better results in your prompt project:

1. Clarify the task : When designing prompts, first clarify what the task is. A clear task description helps the model understand your needs accurately.

2. Provide examples : When necessary, providing relevant examples can help the model better understand the task requirements. Especially in complex scenarios, few-sample prompts can often significantly improve model performance.

3. Step-by-step reasoning : For complex problems, use the thinking chain strategy to break the problem down into multiple intermediate steps, guiding the model to reason step by step and generate results.

4. Concise and clear : The prompt words should be as concise and clear as possible, avoiding complex sentence structures and professional terms so that the model can understand them quickly.

5. Multiple experiments : Different prompt words may produce different output effects, so it is recommended to conduct multiple experiments to observe which prompt words can produce the best results.

6. Conclusion

Prompt word engineering is an emerging discipline. Through reasonable prompt word design, we can effectively improve the output quality of large language models. In practical applications, whether in education, medical care or customer service, prompt word engineering has shown great potential. By gaining a deeper understanding of zero-sample prompts, few-sample prompts and thought chain strategies, we can better utilize this technology and interact more efficiently with intelligent systems.