← 返回首页
AI·ChatGPT第 5 期约 4 分钟

The Prompt Engineer’s First Day

提示词工程师的第一天

显示中文译文

Lena’s new business card said “Prompt Engineer,” a job title that did not exist five years ago. Her mother still introduces her as “something with computers.”

Her first task sounded easy: make the AI write a polite refusal email. She typed, “Say no to this invitation.” The AI produced a two-word reply: “No thanks.” Technically perfect. Socially catastrophic.

Lena quickly learned the first rule of her trade: the AI is brilliant but literal. It does exactly what you say, not what you mean. So she rewrote the prompt, adding context, tone, and an example of a good refusal. This time, the email was warm, firm, and gracious.

By noon, she had iterated eleven times. Her colleague watched her work and laughed. “You’re not programming,” he said. “You’re negotiating.” Lena nodded. “Exactly. And the other side never gets tired.”

At the end of the day, her manager asked how it went. “I spent eight hours talking to a machine,” Lena said, “and the machine taught me to write more clearly than four years of college ever did.” Her manager smiled. “That’s the job.”

本期生词

6 个词 · 点击发音

prompt

/prɒmpt/ n. / v.

提示词;提示、促使

A good prompt includes the context, the tone, and an example.

好的提示词要包含背景、语气和示例。

literal

/ˈlɪtərəl/ adj.

字面意思的;按字面的(AI 常犯的错误)

The AI is very literal — it does what you say, not what you mean.

AI 很字面化——你说什么它做什么,而不是领会你的意思。

context

/ˈkɒntekst/ n.

上下文;背景信息

Without context, even a smart model gives generic answers.

没有上下文,再聪明的模型也只能给出泛泛的回答。

iterate

/ˈɪtəreɪt/ v.

迭代;反复改进

We iterated on the design six times before launch.

上线前,这个设计我们迭代了六版。

hallucinate

/həˈluːsɪneɪt/ v.

(AI)产生幻觉,一本正经地编造事实

The model hallucinated a citation that doesn’t exist.

模型编造了一条根本不存在的引用。

guardrail

/ˈɡɑːdreɪl/ n.

护栏;安全限制(防止 AI 出错的规则)

We added guardrails so the chatbot never gives medical advice.

我们加了护栏规则,确保聊天机器人绝不提供医疗建议。