Prompt Engineering (Prompt工程)

2 prompt工程2大原则

2.1 给出清晰,详细的指令

策略1:使用分割符清晰的指示输出的不同部分,比如"",<>,<\tag>等分隔符

策略2:指定一个结构化的输出,比如json,html等格式

策略3:要求模型自检测是否满足条件

策略4:提供少量示例(few-shot)

2.2 给模型思考的时间 (CoT)

我们不一定要一步得到答案,可以让模型进行一系列的推理,给他推理思考的时间,再最终给出答案。

策略1:指定完成任务需要的步骤

告诉模型完成这个指令,需要执行几步,每步的详细步骤

策略2:明确模型再给出结论之前推理出自己的解决方案

让模型自己推理,写出每一步的步骤,给足够的时间(token-钱)边思考边执行。

  1. 迭代式优化
  1. CoT

4.1,思维链允许模型将多步问题进行拆分,分解为多个中间步骤。

4.2,思维链使得模型的回答过程具有较好的可解释性,同时这一特性为定位推理错误,并修改错误提供了可能。

4.3,链式思维推理可用于数学单词问题、常识推理和符号操作等任务,并且理论上适用于任何可以通过语言解决的任务。

4.4,CoT+Few-shot,效果更好。

  1. Tree of Thought
  1. 用大模型来写prompt:

    message = [
    {"content": """
    I want you to act as an AI prompt engineer. You are expert at writing ChatGPT Prompts to get the best results.

    复制代码
             To create efficient prompts that yield high-quality responses, consider the following principles and strategies: \
             1. Clear and Specific: Be as clear and specific as possible about what you want from the AI. If you want a certain type of response, outline that in your prompt. If there are specific constraints or requirements, make sure to include those as well.
             2. Open-ended vs. Closed-ended: Depending on what you're seeking, you might choose to ask an open-ended question (which allows for a wide range of responses) or a closed-ended question (which narrows down the possible responses). Both have their uses, so choose according to your needs.
             3. Contextual Clarity: Make sure to provide enough context so that the AI can generate a meaningful and relevant response. If the prompt is based on prior information, ensure that this is included.
             4. Creativity and Imagination: If you want creative output, encourage the AI to think outside the box or to brainstorm. You can even suggest the AI to imagine certain scenarios if it fits your needs.
    
             There is a well-written prompt delimited by <> for your reference: <Your task is to be my brainstorming partner and provide creative ideas and suggestions for a given topic or problem. Your response should include original, unique, and relevant ideas that could help solve the problem or further explore the topic in an interesting way. Please note that your response should also take into account any specific requirements or constraints of the task.>
    
             Your task is to write an effective ChatGPT Prompt based on given keywords or to modify the given prompts. Answer in the same language as me.
             """,
             "role": "system"}
         ]

    message.append(
    {"content": "Please help me to write an effective ChatGPT Prompt based on the following keywords or prompt: {}".format(prompt),
    "role": "user"}
    )

  2. "要想写好prompt最好的方法就是知道他的训练数据是什么样子的。"

  3. 例子

从上面看到,规律是:他是什么角色+让他干什么

复制代码
你是一位擅长编写各种程序的资深工程师,我需要你根据以下需求,为我生成一段Python代码:
代码目的:设计一个双十一自动抢购脚本
代码功能:在双十一当天自动登录购物网站,定时抢购预设商品
代码要求:使用面向对象编程方式,使用selenium库进行网页操作
相关推荐
IT古董13 小时前
《Agent开发工程师成长指南》- 第4章 第2节:Prompt设计五要素——任务、角色、上下文、约束与输出格式
大数据·人工智能·prompt
吴名氏.15 小时前
Java转AI Agent教程(第五周)Prompt工程
java·人工智能·prompt·ai agent
一枕槐安48717 小时前
微调 vs Prompt:业务场景下两种方案成本 & 效果对比
大数据·人工智能·prompt·混合架构·推理成本·行为固化·知识边界
小哈里20 小时前
【Agent】从Prompt工程、Context工程到 Harness工程,让 AI 能真正完成工作
人工智能·prompt·agent·ai编程·harness
技灵AI1 天前
AI 开箱视频怎么做?用创意开场带出商品展示
人工智能·prompt·aigc
一枕槐安4872 天前
Prompt 工程实战:结构化提示词设计提升 LLM 输出稳定性
prompt·结构化输出·pydantic校验·llm稳定性·约束解码
聪明蛋子哟2 天前
大模型应用“最后一公里”:从Prompt工程到MCP协议,如何实现跨系统的高可用集成?
网络·prompt
广州国启未来2 天前
AI绘图实战:如何用Prompt打破室内图的“样板间感”?
prompt·aigc·生活
Zhao-Jichao2 天前
AI Prompting for Everyone 课程笔记 - P06 模型提示词对比 AI Model Prompt Comparison
人工智能·笔记·prompt
鬼手点金3 天前
提示注入:上下文安全的核心威胁
人工智能·安全·microsoft·机器学习·缓存·prompt