AI时代，有效的提示工程成为关键技能。AI初创公司Anthropic近期发布了《提示工程概述》，旨在指导用户如何与AI模型（特别是其Claude模型）进行高效交互。指南建议将Claude视为“一位博学但失忆的新员工”，需要清晰、具体的指令。文章强调了提供明确的背景信息、目标受众和最终目的的重要性，并建议使用列表形式组织指令。此外，通过提供示例（多轮提示）、允许AI“思考”（思维链提示）以及赋予AI特定角色（角色提示）等策略，可以显著提高AI的响应准确性和质量。为减少AI“幻觉”，建议允许AI承认不确定性，并要求其引用来源进行核实。

💡 **明确指令与背景信息是关键**：与AI交互时，应将其视为一个需要清晰指导的新手。提供具体的任务目标、预期的结果、目标受众以及工作流程，能够显著提升AI的理解和响应准确性。将指令以列表形式呈现，有助于AI更有效地处理复杂请求。

🌟 **利用示例提升输出质量**：提供精心设计的范例（多轮提示）是引导AI生成精确、一致且高质量内容的有效方法。通过示例，AI能更好地理解用户的期望，减少误解，并遵循统一的结构和风格。

🧠 **赋予AI“思考”空间**：采用“思维链”（Chain of Thought, CoT）提示技巧，鼓励AI逐步分解问题并进行逻辑推理，可以大幅提升其解决复杂问题的能力和输出结果的准确性及深度。

🎭 **角色扮演增强交互精准度**：为AI分配特定角色（如“新闻编辑”或“财务规划师”）是一种强大的提示策略。这有助于AI以特定的语气、风格和专业视角来回应，从而更精准地满足用户在特定场景下的需求。

🚫 **减少“幻觉”的策略**：为了避免AI生成不准确或虚假信息，应明确允许AI承认“不知道”并提供“我不知道”的选项。同时，要求AI引用信息来源进行核实，并在无法找到支持性证据时修正或撤回其陈述，能有效减少“幻觉”的发生。

Writing good prompts is one of the most important skills of the AI era.Anthropic just released a guide on how best to do it.One tip is to think of its chatbot, Claude, as "a brilliant but very new employee (with amnesia)."

One of the most lucrative corporate skills these days is effective prompting — making requests of chatbots to get what you want — but it's not as easy as it sounds.

Chatbots are like children. They will do only what you ask, if even that, and nothing more. To get what you want, you have to be explicit and specific.

AI startup Anthropic is offering some help on this front. The company recently published a "Prompt Engineering Overview" to help users get started.

While its guide applies to pretty much any chatbot, it's tailored to its own, Claude.

The first order of business, Anthropic says, is to understand exactly what Claude is. "When interacting with Claude, think of it as a brilliant but very new employee (with amnesia) who needs explicit instructions," the company says in its guide.

The second is to have a rough idea or draft of your question and a sense of what a successful outcome might look like. Anthropic also offers a "prompt generator" for the first draft.

Then it becomes all about refining that initial prompt. Here are Anthropic's top tips.

Be specific with your prompts

"Claude does not have context on your norms, styles, guidelines, or preferred ways of working. The more precisely you explain what you want, the better Claude's response will be," Anthropic says.

The company suggests telling the chatbot what the results will be used for and what audience it is meant for. You should also tell Claude, or whatever chatbot you are using, what the end goal of the task is.

The more you can organize the directions, the better. Anthropic even recommends laying out the requests as bullet points or a numbered list.

Be generous with examples

"Examples are your secret weapon shortcut for getting Claude to generate exactly what you need," Anthropic says. "By providing a few well-crafted examples in your prompt, you can dramatically improve the accuracy, consistency, and quality of Claude's outputs."

This strategy is sometimes called multi-shot prompting. Anthropic says that giving examples reduces misinterpretation and enforces uniform structure and style.

Give the chatbot space to think

"Giving Claude space to think can dramatically improve its performance," Anthropic says. "This technique, known as chain of thought (CoT) prompting, encourages Claude to break down problems step-by-step, leading to more accurate and nuanced outputs."

This means a user will get the most out of a chatbot if they lay out the chain of steps so it can think through each one before answering. "This thorough reasoning leads to a more confident and justifiable recommendation," Anthropic says.

Roleplay

Anthropic says one of the most effective strategies is to assign the chatbot a specific role, like "news editor" or "financial planner."

"This technique, known as role prompting, is the most powerful way to use system prompts with Claude," the company says.

"In complex scenarios like legal analysis or financial modeling, role prompting can significantly boost Claude's performance."

Assigning roles can ensure you get exactly what you want. Maybe you want the brevity of a news writer, or maybe you want the tone of an academic.

Reduce hallucinations

Chatbots make things up, which is why you have to check everything they say. But there are some simple ways to reduce those hallucinations.

Anthtropic says the best thing you can do is give the chatbot permission to say, "I don't know."

"Explicitly give Claude permission to admit uncertainty. This simple technique can drastically reduce false information," Anthropic says.

You can also ask Claude and other chatbots to cite their claims with sources. "You can also have Claude verify each claim by finding a supporting quote after it generates a response. If it can't find a quote, it must retract the claim," Anthropic says.