Prompt Engineering For ChatGPT: A Quick Guide To Techniques, Tips, And Best Practices
ChatGPT 的提示工程：技巧、建议和最佳实践的快速指南

PROMPT ENGINEERING FOR CHATGPT
聊天 GPT 的提示工程

A Quick Guide To Techniques, Tips, And Best Practices
快速指南：技巧、提示和最佳实践

Contents 内容

1 Introduction 1 引言

1.1 Brief overview of ChatGPT
1.1 ChatGPT 简要概述

1.2 Importance of prompt engineering in maximizing the effectiveness of ChatGPT
1.2 提升 ChatGPT 效果的提示工程重要性

1.3 Objective and structure of the article
1.3 文章的目标和结构

1. Fundamentals of Prompt Engineering
   提示工程基础
2. Techniques for Effective Prompt Engineering
   有效提示工程的技术
3. Best Practices for Prompt Engineering
   提示工程的最佳实践
4. Advanced Prompt Engineering Strategies
   高级提示工程策略
5. Case Studies: Real-World Applications of Prompt Engineering
   案例研究：提示工程的实际应用
6. Conclusion 结论

2 Fundamentals of Prompt Engineering
提示工程的两个基本原则

2.1 What is prompt engineering?
2.1 什么是提示工程？

2.2 The role of prompts in interacting with ChatGPT
2.2 提示在与 ChatGPT 互动中的作用

Response 1: 响应 1：

Response 2: 响应 2：

2.3 Factors influencing prompt selection
2.3 影响提示选择的因素

• User intent: Understand the user’s goal and desired output. This helps in crafting a prompt that aligns with the user’s expectations. Consider the purpose of the interaction, whether it’s for information retrieval, content generation, or problem-solving.
  用户意图：理解用户的目标和期望的输出。这有助于制定与用户期望一致的提示。考虑互动的目的，无论是信息检索、内容生成还是问题解决。
• Model understanding: Familiarize yourself with the strengths and limitations of ChatGPT. This knowledge assists in designing prompts that exploit the model’s capabilities while mitigating its weaknesses. Keep in mind that even state-of-the-art models like ChatGPT may struggle with certain tasks or produce incorrect information.
  模型理解：熟悉 ChatGPT 的优点和局限性。这些知识有助于设计利用模型能力的提示，同时减轻其弱点。请记住，即使是最先进的模型如 ChatGPT 也可能在某些任务上遇到困难或产生不正确的信息。
• Domain specificity: When dealing with a specialized domain, consider using domain-specific vocabulary or context to guide the model towards the desired response. Providing additional context or examples can help the model generate more accurate and relevant outputs.
  领域特异性：在处理专业领域时，考虑使用特定领域的词汇或上下文来引导模型朝着期望的响应方向发展。提供额外的上下文或示例可以帮助模型生成更准确和相关的输出。
• Clarity and specificity: Ensure the prompt is clear and specific to avoid ambiguity or confusion, which can result in suboptimal responses. Ambiguity can arise from unclear instructions, vague questions, or insufficient context.
  清晰和具体性：确保提示清晰具体，以避免模糊或混淆，这可能导致次优的回应。模糊性可能源于不明确的指示、模糊的问题或不足的上下文。
• Constraints: Determine if any constraints (e.g., response length or format) are necessary to achieve the desired output. Explicitly specifying constraints can help guide the model towards generating responses that meet specific requirements, such as character limits or structured formats.
  约束：确定是否需要任何约束（例如，响应长度或格式）以实现所需的输出。明确指定约束可以帮助引导模型生成符合特定要求的响应，例如字符限制或结构化格式。

3 Techniques for Effective Prompt Engineering
有效提示工程的三种技术

3.1 Clear and specific instructions
3.1 清晰具体的指示

Example: 示例：

3.2 Using explicit constraints
3.2 使用显式约束

Example: 示例：

3.3 Experimenting with context and examples
3.3 实验上下文和示例

Example: 示例：

3.4 Leveraging System 1 and System 2 questions
3.4 利用系统 1 和系统 2 的问题

Response: "The capital of France is Paris."
法国的首都是巴黎。

3.5 Controlling output verbosity
3.5 控制输出详细程度

Example: 示例：

• Evaporation: Solar energy heats water in oceans, lakes, and rivers, turning it into water vapor that rises into the atmosphere.
  蒸发：太阳能加热海洋、湖泊和河流中的水，将其转化为上升到大气中的水蒸气。
• Transpiration: Plants release water vapor into the air through small openings in their leaves, called stomata.
  蒸腾作用：植物通过叶子上的小孔（称为气孔）将水蒸气释放到空气中。
• Condensation: As water vapor rises and cools, it condenses into tiny water droplets or ice crystals, forming clouds.
  凝结：当水蒸气上升并冷却时，它会凝结成微小的水滴或冰晶，形成云。
• Precipitation: When cloud particles grow large enough, they fall back to Earth as rain, snow, hail, or sleet.
  降水：当云粒子长大到足够大时，它们会以雨、雪、冰雹或冻雨的形式落回地球。
• Runoff: Precipitation that reaches the ground flows into streams, rivers, and eventually back into oceans, where the cycle begins anew.
  径流：降水到达地面后流入溪流、河流，最终回到海洋，循环重新开始。
• Infiltration: Some of the precipitation seeps into the ground, replenishing underground water reserves called aquifers.
  渗透：部分降水渗入地下，补充称为含水层的地下水储备。

4 Best Practices for Prompt Engineering
提示工程的四个最佳实践

4.1 Iterative testing and refining
4.1 迭代测试与完善

Example: 示例：

4.2 Balancing user intent and model creativity
4.2 平衡用户意图与模型创造力

Example: 示例：

4.3 Harnessing external resources and APIs
4.3 利用外部资源和 API

import openai_secret_manager
import requests

api_key = openai_secret_manager.get_secret("OpenWeatherMap")["api_key"]
location = "San Francisco, US"
url = f"http://api.openweathermap.org/data/2.5/weather?q={location}&appid={api_key
    }&units=metric"
response = requests.get(url)
weather_data = response.json()
temperature = weather_data["main"]["temp"]
weather_description = weather_data["weather"][0]["description"]
prompt = f"The current weather in {location} is: {weather_description}. The
    temperature is {temperature} degrees Celsius. Can you provide a brief summary
    of the weather?"
chatgpt_response = chatgpt.generate(prompt)

import wikipediaapi
wiki = wikipediaapi.Wikipedia("en")
page_title = "Natural language processing"
page = wiki.page(page_title)
summary = page.summary[0:500]
prompt = f"The Wikipedia summary of {page_title} is:\n{summary}\nCan you provide a
    concise explanation of natural language processing in your own words?"
chatgpt_response = chatgpt.generate(prompt)

4.4 ChatGPT OpenAI API example
4.4 ChatGPT OpenAI API 示例

import openai
openai.api_key = "your_openai_api_key_here"
def chat_with_gpt(prompt):
response = openai.Completion.create(
engine="text-davinci-002",
prompt=prompt,
max_tokens=50,
n=1,
stop=None,
temperature = 0.8,
)
return response.choices[0].text.strip()
prompt = "Write a brief introduction to the history of computers."
response_text = chat_with_gpt(prompt)

print(response_text)

• Importing required libraries: We start by importing the openai library, which is the official Python library for OpenAI’s API.
  导入所需的库：我们首先导入 openai 库，这是 OpenAI API 的官方 Python 库。
• Setting the API key: 设置 API 密钥：

• Defining the chat_with_gpt function:
  定义 chat_with_gpt 函数：

• API call: API 调用：

• engine: The ID of the GPT model. In our example, we use the “text-davinci-002” engine.
  引擎：GPT 模型的 ID。在我们的例子中，我们使用“text-davinci-002”引擎。
• prompt: The text prompt that we want the model to respond to.
  提示：我们希望模型响应的文本提示。
• max_tokens: The maximum number of tokens (words or word pieces) in the generated response.
  最大令牌数：生成响应中令牌（单词或单词片段）的最大数量。
• n: The number of generated responses.
  生成的响应数量。
• stop: An optional sequence that indicates the end of a response.
  停止：一个可选的序列，表示响应的结束。
• temperature: Controls the randomness of the output. A higher value makes the output more random, while a lower value makes it more deterministic.
  温度：控制输出的随机性。较高的值使输出更随机，而较低的值则使其更确定。
• Processing the response: 处理响应：

- Using the function: - 使用该功能：

• Printing the response: Finally, we print the generated response to the console using the print() function.
  打印响应：最后，我们使用 print() 函数将生成的响应打印到控制台。

4.5 Ensuring ethical usage and avoiding biases
4.5 确保伦理使用和避免偏见

1. Being aware of potential biases: Familiarize yourself with the possible biases that may arise in ChatGPT’s responses. This awareness will help you identify and address such biases when crafting prompts.
   意识到潜在偏见：熟悉可能在 ChatGPT 的回答中出现的偏见。这种意识将帮助您在撰写提示时识别和解决这些偏见。
2. Using inclusive language: When designing prompts, use language that encourages diverse perspectives and avoids reinforcing stereotypes. This approach helps ensure that the generated content is inclusive and respectful.
   使用包容性语言：在设计提示时，使用鼓励多样化观点的语言，避免强化刻板印象。这种方法有助于确保生成的内容具有包容性和尊重性。
3. Evaluating generated content: Regularly assess the content generated by ChatGPT for potential biases or ethical concerns. If you discover issues, refine the prompts to mitigate them.
   评估生成的内容：定期评估 ChatGPT 生成的内容，以发现潜在的偏见或伦理问题。如果发现问题，请调整提示以减轻这些问题。
4. Implementing content filters: Utilize content filters or moderation tools to screen the generated responses for potentially harmful or biased content before presenting them to users.
   实施内容过滤器：利用内容过滤器或审核工具在向用户展示生成的回复之前，筛选出潜在有害或偏见的内容。

Example: 示例：

5 Advanced Prompt Engineering Strategies
5 种高级提示工程策略

5.1 Temperature and token control
5.1 温度和令牌控制

• Prompt: “Write a short poem about spring.”
  春天的诗 春风轻拂花开时， 万物复苏绿意浓。 小鸟欢唱枝头舞， 阳光洒下暖融融。
• Response: 响应：
  “Spring has arrived with a gentle breeze, Awakening flowers and rustling leaves. The sun shines bright, the days grow long, Birds sing their sweet and joyful song.”
  春天伴随着轻柔的微风到来了，唤醒了花朵和沙沙作响的树叶。阳光明媚，日子渐渐变长，鸟儿唱着甜美而快乐的歌。

• Prompt: “Write a short poem about spring.”
  春天的诗 春风轻拂花开时， 万物复苏绿意浓。 小鸟欢唱枝头舞， 阳光洒下暖融融。
• Response: 响应：
  “Beneath the sun’s glistening crown,
  在阳光闪耀的冠冕下，
  The once-silent world starts to resound.
  曾经寂静的世界开始回响。
  Spring is a vibrant dancer on the stage,
  春天是舞台上充满活力的舞者，
  Unfurling colors in a whimsical parade.”
  在奇幻的游行中展开色彩。

5.2 Prompt chaining and multi-turn conversations
5.2 提示链和多轮对话

Example: 示例：

5.3 Adapting prompts for domain-specific applications
5.3 针对特定领域应用调整提示

Example: 示例：

5.4 Handling ambiguous or contradictory user inputs
5.4 处理模糊或矛盾的用户输入

Example: 示例：

6 Case Studies: Real-World Applications of Prompt Engineering
6 个案例研究：提示工程的实际应用

6.1 Customer support chatbots
6.1 客户支持聊天机器人

Example: 示例：

1. Go to the account login page.
   前往账户登录页面。
2. Click on the ‘Forgot Password’ link.
   点击“忘记密码”链接。
3. Enter the user’s registered email address.
   输入用户注册的电子邮件地址。
4. Click ‘Submit’ to send a password reset link to the user’s email.
   点击“提交”以将密码重置链接发送到用户的电子邮件。
5. Ask the user to check their email and click on the link provided.
   请用户检查他们的电子邮件并点击提供的链接。
6. They will be redirected to a password reset page, where they can create a new password.
   他们将被重定向到密码重置页面，在那里可以创建新密码。
7. Once the new password is set, instruct the user to $\log$$\log$log\log in using their new credentials."
   一旦新密码设置完成，请指导用户 $\log$$\log$log\log 使用他们的新凭据。

6.2 Content generation and editing
6.2 内容生成与编辑

Example: 示例：

6.3 Domain-specific knowledge retrieval
6.3 领域特定知识检索

Example: 示例：

1. Marigolds: These vibrant flowers are low-maintenance and thrive in a variety of soil types.
   金盏菊：这些色彩鲜艳的花朵易于养护，适应多种土壤类型。
2. Basil: This popular herb grows quickly and can be used in a wide range of dishes.
   罗勒：这种受欢迎的草药生长迅速，可以用于多种菜肴。
3. Tomatoes: Choose a hardy variety like cherry tomatoes for a bountiful harvest.
   番茄：选择像樱桃番茄这样的耐寒品种，以获得丰收。
4. Zucchini: These fast-growing vegetables are perfect for garden beds or containers.
   西葫芦：这些生长迅速的蔬菜非常适合花园床或容器种植。
5. Sunflowers: With their cheerful appearance, sunflowers are easy to grow and require minimal care.
   向日葵：向日葵以其愉悦的外观而闻名，易于种植且需要的护理很少。
6. Remember to provide adequate sunlight, water, and well-draining soil for optimal growth."
   记得提供充足的阳光、水分和良好的排水土壤，以促进最佳生长。

6.4 Interactive storytelling and gaming
6.4 互动故事讲述与游戏

Example: 示例：

7 Conclusion 7 结论

7.1 The impact of effective prompt engineering on ChatGPT performance
7.1 有效提示工程对 ChatGPT 性能的影响

7.2 Future directions in prompt engineering research and applications
7.2 提示工程研究和应用的未来方向

7.3 Encouraging creativity and collaboration in the ChatGPT community
7.3 鼓励 ChatGPT 社区中的创造力和合作

References 参考文献