新智元报道
提示词工程，死了。
给它判死刑的人，正是当年亲手把「提示词工程师」变成一份正式工作的那个男人——Riley Goodside。
祖师爷，亲手给自己的饭碗写了一篇讣告。
事情是这样的。
10 月 22 日，安全研究员 Justin Elze 在 X 上发了一条帖子，正文只有一句：You're all prompting wrong，你们提示的方式全都错了。
配图是两张 LinkedIn 截图。
第二天，Riley Goodside 转发了它，轻飘飘丢出三句话：
Prompt engineering is dead. Prompt however you want. Have fun。提示工程已死，想怎么问就怎么问，玩得开心。
三句话，加起来 10 个英文单词。可就这 10 个词，在 AI 圈掀起了滔天巨浪。
这话不是谁说都有这分量。
Goodside 做过 Scale AI 的首席提示词工程师，后来又加入 Google DeepMind 继续干这行。Scale AI 的 CEO Alexandr Wang当年亲自站台说，他打赌 Goodside 是全球第一个被正式雇用的提示词工程师。
早年他靠公开做实验出名，专门研究怎么跟大模型打交道，连「提示注入攻击」这个安全概念都是他早期研究的重点方向之一。
有网友评论一针见血：这话出自一个几乎亲手设立了这个职位的人之口，真是一篇有趣的悼词。
把时钟往回拨一拨。
ChatGPT 刚火那阵子，谁的收藏夹里没躺过几份「万能提示词模板」呢？
先给 AI 指定一个身份——「你是一位资深 XX 专家」，再规定输出格式，再叮嘱一句「请一步步思考」，最后还得加两句客气话。
一长串写下来，跟念咒似的，生怕漏掉一个字 AI 就不灵了。
那时候大家是真信这个。
会写提示词，就是一门能写进简历的技能，「提示词工程师」一度是全网最让人眼红的新职业。
据说年薪能开到几十万美元，不用写代码，光靠跟 AI 聊天就行。
三年过去，发明这套「咒语」的人自己站出来说：别念了，没用了。
Justin 配的那两张截图就是活生生的证明。
截图来自 Chainguard 的 CEO Dan Lorenc。
这家公司在软件供应链安全领域赫赫有名。
就在 10 月中旬，Chainguard 正式推出了他们的 AI 漏洞发现系统——一套自建的执行框架，配上前沿大模型做安全研究。
给模型的提示简单到什么程度呢？大意就是一句话：找找 PHP 里不用登录就能被远程利用的高危漏洞。
没有角色扮演，没有分步模板，也没有「你是一位资深安全专家，请按以下步骤……」。
就这么一句大白话。
结果那套系统真的独立复现出了已知的高危漏洞，还帮忙验证了官方修复方案。
这事闹得可动静不小。FBI 都发了警告，PHP 基金会紧急发布修复，多个政府网站一度受影响。
搁在以前，这是资深安全专家反复打磨提示词、精心设计多轮对话才敢碰的活儿。
如今一句人话就搞定了。
热闹没过多久，Goodside 自己在评论区补上了后半句。
他说自己刚才是在说俏皮话。
给系统写底层提示词的人、给小模型做精调的人、把提示嵌进应用里的工程师，手里的活可一点没少。但对坐在屏幕前的普通用户来说，「直接问」能达到的效果上限，基本已经没有天花板了。
有人追问，是哪个新模型让你突然改了想法？
他说没有什么突然的转折，自从 ChatGPT 问世起，提示词工程作为普通用户的必备技能就一直在慢慢消亡，这是个缓慢而持续的过程。
有网友一语道破：「感觉工作只是从提示转移到了 harness 上。」Goodside 回了两个字：是的。
这就对了。
Dan Lorenc 那句「找找漏洞」之所以能成，靠的不是这句话本身有多精妙，而是它背后那套包含沙箱环境、目标源码和执行引擎的完整系统。
模型的功劳在能力，系统的功劳在设计。
把这几年的演化串起来，脉络清晰得很：
最早是 Prompt Engineering，琢磨怎么把一句话说漂亮；后来 Karpathy 等人推动了 Context Engineering，研究怎么给模型喂对信息；再到 Harness Engineering，给模型搭好运行环境、设定规则和约束；现在圈子里又喊出了 Loop Engineering，让整套系统自己循环着跑、自我迭代。
一层一层往上叠，功夫离提示词越来越远。
普通用户不用再背咒语、存模板，模型听得懂人话了。
最有意思的是 Goodside 本人。有人问他，现在该把精力花在什么技能上。这位祖师爷很老实地回答：老实说，我完全不知道现在什么技能值得培养，一切变得太快了。
技巧每隔几个月就被模型追平一次。
现在变得重要的三件事成了：你能不能把问题想清楚？能不能把自己要什么说明白？能不能判断 AI 给出来的结果到底对不对？
参考资料：https://x.com/goodside/status/2113639196254671259
编辑：所罗门
不喜欢
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
New Ji-won.
prompt Project, dead.
The man who sentenced him to death was the man who himself turned "Engineer prompt" into an official job, Riley Goodside.
Master, he wrote an obituary for his own job.
Here's the thing.
On October 22, Security Researcher Justin Elze posted a post on X, with only one sentence: You're all moving wrong.
Two LinkedIn screenshots.
The next day, Riley Goodside relayed it, drifting out three sentences:
Prompt engagement is dead. Tell the project to die. Ask as much as you like. Have fun.
Three sentences, plus 10 words in English. But for the 10 words, there was a giant wave in the AI Circle.
That's not what anyone says.
Goodside was the chief prompt engineer for Scale AI and later joined Google DeepMind to continue the business. The CEO Alexander Wang of Scale AI himself said that he was betting that Goodside was the first prompt engineer in the world to be officially employed.
In his early years, he was famous for his public experiments, his special study of how to deal with LLM, and even the concept of "inflicted attack" as a safety concept was one of his early priorities.
There's an online comment on blood: it's an interesting eulogy that comes from a man who created the post almost personally.
Call the clock back.
ChatGPT, who didn't lie in a couple of "Mighty prompt Templates" in his collection during the fire?
First, you assign AI an identity — “You're a senior XX expert” — then you specify the output format, then you say, “Please think one step at a time”, and then you add two more words.
It's a long line, like a spell, and I'm afraid I'm losing one word.
People believed that.
prompt, which was once one of the most popular new occupations on the Internet, is the skill of writing a curriculum vitae.
They say it's worth hundreds of thousands of dollars a year without writing codes, just talking to AI.
Three years later, the man who invented this spell stood up and said, "Stop it, it's useless."
Justin's two screenshots are living proof.
Image by CEO Dan Lorenc from Chainguard.
This company is well known in the area of software supply chain security.
Just in mid-October, Chainguard officially launched their AI Gap Discovery System, a self-built implementation framework with front-line LLM for safety research.
How simple is the clue to the model? To the effect that is to say, look for high-risk holes in PHP that can be used remotely without login.
There is no role-playing, no step-by-step template and no "you are a senior security expert, please follow the following steps".
That's the big white word.
As a result, the system had effectively independently re-emerged known high-risk gaps and had also helped validate official rehabilitation programmes.
This is quite a scene. The FBI issued warnings that the PHP Foundation had emergency publish repairs and that several government websites had been affected.
Previously, this was a job that senior security experts had repeatedly sharpened prompt and carefully designed multiple rounds of dialogue to take on.
Now a word is done.
Not long ago, Goodside himself added the last part to the comment section.
He said he was talking funny.
The people who write the bottom prompt to the system, the people who make fine models, the engineers who embed the clues in the application, have very little to do. But for ordinary users sitting in front of the screen, there is almost no ceiling left for "direct questions".
I'm being asked, which new model did you suddenly change your mind?
He said that there had been no sudden reversals and that it was a slow and continuous process that prompt had been slowly disappearing since the advent of ChatGPT as an essential skill for ordinary users.
"It's just a feeling that work is moving from tips to papers." Goodside returned two words: yes.
That's right.
Dan Lorenc's "seeing the holes" is not about how fine it is in itself, but about the complete system behind it, which contains the sandbox environment, the target source code and the implementation engine.
The model is capable and the system designed.
It's been a few years, and it's clear:
The first was Prompt Engineering, thinking about how to say a good word; later Karpathy and others pushed Context Engineering to study how to feed the models; then Harness Engineering, putting the models in place, setting rules and constraints; and now the Loop Engineering is in the circle, making the whole system run around and interact.
One layer is stacked up, and kung fu is getting farther from prompt.
Ordinary users don't have to re-enact the spell, store the template, and the model understands the human language.
Most interesting is Goodside himself. He was asked what skills it was time to focus on. The Grand Master replied: I honestly have no idea what skills are worth developing now, and everything is getting too fast.
The technique is pursued every few months.
Three things are important now: can you think of something? Can you explain what you want? Can you judge if AI gave the right results?
References: https://x.com/goodside/status/2136339196254671259
Editor: Solomon
I don't like it.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.