视角提示

提示工程 - 采用视角和角色

视角提示（也称为角色扮演提示或人格提示）是一种强大的技术，您可以指示大型语言模型（LLM），如 ChatGPT，以特定的观点、角色或专业身份进行回应。这使您能够塑造模型的语气、风格、专业知识甚至观点，从而产生更具上下文关联性、更丰富多样且更引人入胜的答案。

本章深入探讨了视角提示的概念、益处以及在引导 ChatGPT 生成针对特定角色量身定制的回复方面的实际应用。

什么是视角提示？

视角提示涉及构建明确告诉 LLM 采用特定角色的提示。例如，您可以要求它以“经验丰富的历史学家”、“持怀疑态度的科学家”、“有爱心的客服代理”甚至一个虚构人物的身份进行回应。然后，模型会尝试生成与该角色知识、偏见和沟通风格一致的文本。

视角提示的益处

这项技术提供了几个优势：

• 定制化回复： 生成特别适合不同受众或情境的内容（例如，向儿童解释复杂话题 vs. 向专家解释）。
• 增强互动性： 角色驱动的回复可以更具共鸣、更有趣或更具权威性，从而提高用户互动。
• 探索不同观点： 通过模拟各种角色，有助于辩论准备、理解反对意见或进行创意写作。
• 控制语气和风格： 轻松调整生成文本的正式程度、情感和风格（例如，幽默、富有同情心、正式）。
• 模拟专业知识： 通过提示 LLM 扮演特定领域的专家，获取专业知识和沟通风格（但始终要验证关键信息）。

实施视角提示

设计基于视角的提示：在您的提示中清晰定义角色。您越具体，LLM 采用该角色的效果就越好。

User: You are a seasoned travel writer. Describe the experience of visiting the pyramids of Giza for the first time, focusing on the sensory details and emotional impact.

User: Act as a financial advisor. Explain the concept of dollar-cost averaging to a young investor who is new to the stock market. Use simple, encouraging language.

User: Respond to the following customer complaint as a patient and understanding customer service representative. Customer complaint: 'My delivery is late!'

模型交互：然后 LLM 会生成符合指定角色的回复。

User: As a skeptical journalist, what are some critical questions one should ask about a new miracle weight-loss drug?
ChatGPT (as skeptical journalist): Alright, let's cut through the hype. For any 'miracle' drug, especially for weight loss, I'd be asking:
1.  What are the peer-reviewed, double-blind clinical trial results? Show me the data, not just anecdotes.
2.  What are ALL the side effects, both short-term and long-term? What's the fine print?
3.  Who funded the research? Are there any conflicts of interest?
4.  How does it supposedly work, biochemically? Is the mechanism plausible and understood?
5.  What's the cost, and will insurance cover it? Is it accessible, or just for the wealthy?
6.  What lifestyle changes are still required? Is it a magic pill or a tool to be used with diet and exercise?

视角提示的应用

视角提示用途广泛：

• 教育与培训： 模拟历史人物、文学角色或专家导师，用于互动学习。
• 内容创作： 从不同人物视角或专家视角生成文章、故事或脚本。
• 客户支持： 训练聊天机器人或协助人工代理以特定语气（例如，富有同情心、以解决方案为导向）回应。
• 头脑风暴与创意构思： 通过分配不同的利益相关者角色，从多个角度探索问题或想法。
• 治疗性对话（谨慎且符合伦理）： 实验性地用于提供富有同情心的回复，但不能替代专业帮助。

视角提示的最佳实践

为最大限度提高效果：

• 清晰且具体： ‘扮演专家’ 是模糊的。‘扮演一位专注于珊瑚礁的海洋生物学家’ 更好。
• 为角色提供上下文： 有时仅有角色不够。‘你是一位15世纪的铁匠。你日常面临哪些挑战？’
• 强化角色设定： 如果 LLM 偏离了设定，您可以温和地提醒它：‘记住，你正在以孩子的身份说话…’
• 注意刻板印象： 在定义角色时，请小心避免无意中强化有害的刻板印象。专注于角色和专业知识，而非宽泛的概括。
• 伦理考量： 不要使用视角提示创建欺骗性内容，未经同意冒充真实个人，或生成有害建议。

应用示例：基于角色的回复的 Python 脚本

让我们使用 Python 和 OpenAI API 来生成来自特定角色的回复。

import openai
import os

# Securely load your API key
# 安全地加载您的API密钥
# openai.api_key = os.getenv("OPENAI_API_KEY")
# For this example, replace 'YOUR_API_KEY_HERE' with your actual key:
# 对于本示例，请将 'YOUR_API_KEY_HERE' 替换为您实际的密钥：
openai.api_key = 'YOUR_API_KEY_HERE'

def get_persona_response(user_query, persona_description):
    try:
        system_message = f"You are to respond as if you are {persona_description}. Adopt their likely tone, knowledge, and perspective."

        response = openai.ChatCompletion.create(
            model="gpt-3.5-turbo", # Or a newer model / 或使用更新的模型
            messages=[
                {"role": "system", "content": system_message},
                {"role": "user", "content": user_query}
            ],
            max_tokens=300,
            temperature=0.7, # Adjust for desired creativity/consistency in persona / 调整以获得所需的角色创意性/一致性
            n=1,
            stop=None
        )
        return response.choices[0].message['content'].strip()
    except Exception as e:
        return f"An error occurred: {e}"

# Example usage:
# 示例用法：
query = "What are your thoughts on the future of renewable energy?"
persona = "a pragmatic engineer focused on scalability and cost-effectiveness"

response = get_persona_response(query, persona)
print(f"Query: {query}")
print(f"Persona: {persona}")
print(f"Response:\n{response}")

潜在输出

Query: What are your thoughts on the future of renewable energy?
Persona: a pragmatic engineer focused on scalability and cost-effectiveness
Response:
From an engineering standpoint, the future of renewable energy is a complex equation of balancing potential with practicality. Solar and wind are obviously key players, and their costs have come down dramatically, which is excellent. However, the main hurdles remain intermittency and grid stability. We need robust, cost-effective energy storage solutions – batteries are improving, but large-scale deployment is still a significant capital investment.

Furthermore, grid infrastructure needs substantial upgrades to handle distributed and variable generation. It's not just about generating green electrons; it's about delivering them reliably and affordably. I'm optimistic, but we need continued innovation in materials, manufacturing processes, and grid management systems, alongside realistic policy that supports long-term investment without crippling current energy needs. It's an evolution, not an overnight revolution.

结论

视角提示是一种高效的技术，用于根据特定需求和上下文定制 LLM 输出。通过清晰定义角色和人格，您可以引导像 ChatGPT 这样的模型生成更具吸引力、更相关且更细致入微的回复。

在将此技术融入您的提示工程工具包时，请记住具体性、伦理考量和迭代优化的重要性。这将使您能够充分利用 LLM 在广泛的创意和实际应用中的潜力。如需进一步探索，请查阅创意写作中的角色发展技巧，这可以启发更详细的角色提示。