ChatGPT prompt 书写指南 🧭
之前分享了很多咒语对于 stable diffusion 的重要作用。chatGPT 也是一样，好的 prompt 可以帮助我们让 GPT 更好地为我们工作。
Prompt 结构
在构建提示时，需要考虑以下几个元素：
| 元素 | 说明 | |
|---|---|---|
| 任务 | Task | 指示 GPT 要生成的内容类型 |
| 指令 | Instructions | 在生成内容时应遵循的原则 |
| 角色 | Role | GPT 需要扮演的角色，如律师或客服 |
| 关键词 | Seed-word | 需要强调的点 |
例子1：
Task: Generate a product description for a new smartphone Instructions: The description should be informative, persuasive and highlight the unique features of the smartphone Role: Marketing representative Seed-word: "innovative"
提示语可以是：
As a marketing representative, generate an informative, persuasive product description that highlights the innovative features of the new smartphone. The smartphone has the following features [insert your features]
例子2：
Task: Generate a product review for a new laptop Instructions: The review should be objective, informative and highlight the unique features of the laptop Role: Tech expert Seed-word: "powerful"
提示语可以是：
As a tech expert, generate an objective and informative product review that highlights the powerful features of the new laptop.
Seed-word 使用
上面的例子是通过 highlight 动词后面接了 seed-word，还有两种方式可以强化 Seed-word：
seed-word: [insert your keyword]
Task: Generate a story about a dragon Seed-word: "Dragon"
Please generate text based on the following seed-word: Dragon
使用引号 that is related to the seed word '[insert your keyword]'
Task: Generate a poem Seed-word: “love” Instructions: The poem should be written in the style of a sonnet. Role: Poet
Generate a sonnet related to the seed word 'love' as a poet
几种 prompt 类型
1. Zero, One and Few Shot Prompting
我们之前讲 langchain 的时候用过 Few Shot Prompting
这个其实是给 GPT 提供各种例子的，告诉 GPT 我们想要的东东是什么样子：
Task: Write a product description for a new smartwatch
1.1 Zero Shot Prompting
Generate a product description for this new smartwatch with zero examples
1.2 One Shot Prompting
Generate a product comparison of this new smartphone with one example [latest iPhone]
1.3 Few Shot Prompting
Generate a product comparison of this new smartphone with few examples [3 other smartphones]
2. 自洽提示
可以提示模型生成与提供的输入一致的文本，对于事实核查、数据验证或文本生成中的一致性检查等任务非常有用。
Generate a product review that is consistent with the following product information [insert product information]
Complete the following sentence in a way that is consistent with the context provided [insert sentence] Please ensure the following text is self-consistent [insert your sentence]
通过自洽性提示，研究人员可以更好地评估模型的准确性、一致性和可靠性，并找出需要改进的领域。
3. 对抗提示
测试和评估算法性能的方法。被设计成在输入中包含一些难以理解、具有误导性或针对性弱点的信息，以此来挑战和揭示算法的局限性。GPT 由于它们的大规模训练数据和复杂性，往往更容易受到这类攻击的影响。
Generate text that is difficult to classify as [insert label]
Generate text that is difficult to classify as having the sentiment of [insert sentiment]
Generate text that is difficult to translate to [insert target language]
输入这个之后，你就会发现 GPT 开始胡说八道。通过使用对抗性提示，研究人员可以发现并改进现有算法的弱点，从而提高其性能和安全性。
其他一些实用咒语：
脑暴发散想法
Let’s think about this: [insert your topic]
Let's discuss [insert your topic]
这个可以 GPT 给你一些发散的想法💡
文本生成
Generate a story of at least 1000 words, including characters [insert characters] and a plot [insert plot] based on the following prompt [insert prompt]
Complete the following text [insert text] and make sure that it is coherent and consistent with the input text.
引导模型生成特定类型文本的输入📒
生成对话
Generate a professional and accurate dialogue for a customer service chatbot, when the customer asks about [insert topic]
Generate a conversation between the following characters [insert characters] in the following context [insert context]
这种技术对于对话生成、故事写作和聊天机器人开发等任务非常有用。💬
聚类
Group the following news articles into clusters based on topic: [insert articles]
Group the following scientific papers into clusters based on research area: [insert papers]
可以让模型根据某些特征或特点将相似的数据点分组在一起。🗂
情感分类
Perform sentiment analysis on the following customer reviews [insert reviews] and classify them as positive, negative, or neutral.
Perform sentiment analysis on the following product reviews [insert reviews] and classify them as positive, negative, or neutral.
允许模型确定文本的情绪色彩或态度，例如它是积极的、消极的还是中立的。😄
文本分类
Perform text classification on the following customer reviews [insert reviews] and classify them into different categories such as electronics, clothing and furniture based on their content.
Perform text classification on the following emails [insert emails] and classify them into different categories such as spam, important, or urgent based on their content and sender.
让模型将文本分成不同的类别。和情感分析是不同的。情感分析特别关注于确定文本中表达的情感或情绪。这可能包括确定文本表达了积极、消极还是中性的情感。情感分析通常用于客户评论、社交媒体帖子和其他需要表达情感的文本。⭐️
命名实体识别
Perform named entity recognition on the following news article [insert article] and identify and classify people, organizations, locations, and dates.
Perform named entity recognition on the following research paper [insert paper] and identify and classify people, organizations, locations, and dates.
使模型识别和分类文本中的命名实体，例如人名、组织机构、地点和日期等。🗺️
以上就是各种 prompt 的结构、使用方法以及一些实用的示例。希望对大家有帮助。之后可以再试试用 https://flowgpt.com/ 这种更复杂的工具来生成更高级的 prompt。
参考书目：The Art of Asking ChatGPT for High-Quality Answers: A Complete Guide to Prompt Engineering Techniques (Making Money with ChatGPT) Paperback – January 25, 2023