The truly essential skills don't become jobs. They become assumed.
In January 2023, the finance-careers site eFinancialCareers ran a headline engineered to stop scrolling thumbs: "The $335k engineering job where you don't write any code." The job was real. Anthropic, the company behind the Claude models, had posted an opening for a "Prompt Engineer and Librarian" with a listed salary range of $175,000 to $335,000. The duties: "figure out the best methods of prompting our AI to accomplish a wide range of tasks," then "build up a library of tools and a set of tutorials that allows others to learn prompt engineering." By March, Fortune was reporting that someone with a "hacker spirit" could earn over $300,000 in the new role category, and a thousand LinkedIn courses bloomed. Prompt engineering, the story went, was the next six-figure standalone career, a profession for people skilled at talking to machines, in a field that had existed for barely two years.
Read that job posting again, though, and notice what it actually says. Anthropic wasn't just hiring someone to write prompts. It was hiring someone to write the tutorials, to build the library that would teach everyone else. The model-maker was paying up to $335,000 for a priesthood whose explicit mandate was to make the priesthood unnecessary. That detail turned out to be the whole story in miniature. The prediction that prompting would matter enormously was correct. The prediction that mattering enormously would make it a job got the economics exactly backwards.
The reversal arrived with unusual speed and an unusually good punchline. In March 2024, little more than a year after the $335k headline, IEEE Spectrum published a piece titled "AI Prompt Engineering Is Dead." It reported on work by Rick Battle and Teja Gollapudi at VMware, who had done something quietly devastating: instead of having humans hand-tune prompts through trial and error, they asked the model to optimize its own prompts, scored against a benchmark. The result: "in almost every case, this automatically generated prompt did better than the best prompt found through trial and error."
Better is one thing. The weirder part is the one that sticks. One of the top-performing auto-generated prompts for grade-school math problems began by putting the model on the bridge of a starship: "Command, we need you to plot a course through this turbulence and locate the source of the anomaly. Use all available data and your expertise to guide us through this challenging situation." No human prompt engineer would have written that. No human prompt engineer could have defended that in a design review. It won anyway. The machine had found a groove in its own latent space that the professionals couldn't see, which rather undermines the premise that the professionals' intuition was the scarce asset. Spectrum's conclusion was measured but pointed: the findings "increased suspicions that a fair portion of prompt-engineering jobs may be a passing fad, at least as the field is currently imagined."
Hold on to that Star Trek prompt. It's the clearest evidence anyone has produced for a strange claim: the skill at the center of the hottest job of 2023 was one the tool itself was learning faster than the humans hired to have it.
If you want to know how the story ends, you don't need a forecast. You need a memory that reaches back to 1993, the year, per Merriam-Webster, of the first known use of the word webmaster.
The webmaster was the temporary priesthood of the last "arcane new thing." One person who could do HTML and server configuration and graphics and content, keeper of the whole strange machine, at a time when the machine was finicky and the skill was rare. And the role wasn't fringe. It got institutionalized into the internet's actual plumbing. In May 1997, the IETF published RFC 2142, which standardized webmaster@domain as the contact address every website was expected to answer. Think about what that means: the job title was embedded in an internet standard. It's hard to imagine a role looking more permanent.
Then it dissolved, in two directions at once, and this is the part that maps onto 2023 so cleanly it's almost unfair.
It dissolved upward into specialists. Organizations stopped employing "a webmaster" and started employing front-end developers, back-end developers, DevOps engineers, site reliability engineers, SEO specialists, content managers, security and analytics and accessibility people. Every slice of the webmaster's day became somebody's entire career, at salaries the webmaster never saw.
And it dissolved downward into everyone and everything. Dreamweaver, then content-management systems, then cloud platforms absorbed the operational arcana; publishing a webpage went from a specialist act to something a marketing intern does before lunch. A Google user-experience study eventually found that very few web professionals identify as "webmasters" anymore, and in 2020 Google itself retired the word, renaming Google Webmasters Central to Google Search Central. The company that had run the web's biggest gathering place for webmasters concluded the title no longer described anyone. Not because the work vanished, but because the work was now everywhere, split between specialists above and everybody below. There was no middle left for a "webmaster" to stand on.
That's the lifecycle: new tool, then a generalist priesthood at a scarcity premium, then institutional recognition right around the peak, then dissolution up into specialists and down into baseline competence, and finally the title survives only as an email alias nobody reads.
Overlay prompt engineering on it and the joints line up one for one. Hot generalist title at a scarcity premium: 2023, $175k to $335k. Institutional recognition at the peak: the model-maker itself creates the role, with "Librarian" right there in the name. Dissolution upward: "AI engineer," "agent designer," and the term that stuck for the serious tier, more on it in a moment. Dissolution downward: prompting folded into every job description that touches a computer, the way "must be able to use email" stopped being worth typing. The webmaster arc, replayed.
Except for one thing. The webmaster arc took roughly two decades. This one took about two years. And the reason for the speed difference is the genuinely new part of the story.
Here's the asymmetry that makes prompt engineering historically interesting rather than just historically familiar: the webmaster's tool never got better at being operated. This tool did.
An Apache server in 1998 did not study its administrator and start configuring itself. Dreamweaver made humans faster; a human still drove. Every previous tool-operator priesthood, the webmaster, the desktop-publishing specialist, the mainframe computer operator, was eroded from the outside, by other tools and other people gradually absorbing pieces of the work. The operator's core skill stayed human property throughout; it just stopped being scarce.
Prompt engineering is the first case where the tool's own improvement curve pointed directly at the operator's skill. A language model is a machine whose entire product is understanding what you meant, so every capability generation made clever phrasing less necessary. And, per VMware's result, it's also a machine that can generate and test its own instructions, so the residual cleverness could be automated by the very artifact it was cleverness about. The moat wasn't just shallow; the tide coming in was the product roadmap. Even the hiring language of 2023 carried the hedge: what Fortune's coverage of the role emphasized wasn't credentials but a "hacker spirit," an aptitude for fiddling with a brand-new artifact, in a field too young and too fluid for anyone to honestly call it a stable profession.
That's why the half-life was so short. The webmaster was dissolved by the ecosystem. The prompt engineer was dissolved by the tool itself. When the thing you operate is the thing that learns, "operator" is a temporary job category by construction.
Now the necessary correction, because the smug version of this essay ("lol, remember prompt engineers?") gets the ending wrong, and the IEEE piece itself refused to write it. The same article that called the field a passing fad also noted that "prompt-engineering jobs in some form are not going away"; commercial-grade AI products involve a raft of concerns no autotuner touches. Both clauses are true, and the tension resolves the same way the webmaster's did: the middle vanished, the ends thrived.
Look at the up direction. In June 2025, Shopify CEO Tobi Lütke wrote that he preferred a different term for the real skill, "the art of providing all the context for the task to be plausibly solvable by the LLM." Andrej Karpathy amplified it: "+1 for 'context engineering' over 'prompt engineering.' People associate prompts with short task descriptions you'd give an LLM in your day-to-day use. When in every industrial-strength LLM app, context engineering is the delicate art and science of filling the context window with just the right information for the next step." Notice what happened there: the serious tier of the discipline didn't disappear, it got renamed away from the dissolved title, exactly the way "webmaster" became "site reliability engineer." Deciding what a model should see, which documents, which tools, which history, in what order, under what budget, is real engineering, it pays like real engineering, and nobody doing it calls themselves a prompt engineer. The practitioners who insisted "prompt engineering is not dead" were right about the work and wrong about the word; the word was already a fossil.
And the down direction is the punchline the 2023 headlines missed: the skill won so completely that it stopped being distinguishable from working. Writing a decent instruction to a model in 2026 is what using a computer looks like, for lawyers, marketers, support reps, and every developer alive. We build agent systems for a living, and prompting stopped being anyone's title here the day it became part of everyone's job; the entire premise of our work assumes the baseline. "Prompt engineer" in 2026 sounds the way "email specialist" would have sounded in 2005. The universality the hype correctly predicted is precisely what killed the standalone title it incorrectly predicted. Ubiquity doesn't create professions. Ubiquity creates literacy, and nobody gets paid $335,000 for being literate.
(There's a general lesson about hype's chemistry, why we reliably mistake "new and important" for "new job title," but we've written that one separately; see "The Neurochemistry of Hype". And for the companion case of an AI-and-jobs number that was widely misread in the other direction, see "300 Million Jobs and Counting (Still)".)
So here's the practical residue, because this pattern is going to repeat. It may be repeating right now in whatever role your company just posted with "agent" in the title.
When a new tool spawns a new job, ask three questions about the job's center of gravity:
The 2022 prophets weren't wrong that talking to machines would become one of the most valuable skills in the economy. It did, faster than they said. They were wrong about what "most valuable" does to a skill. The truly essential skills don't become jobs. They become assumed. The job title was never the prize; it was the scaffolding, and scaffolding comes down precisely when the building can stand.
Somewhere at Anthropic there's presumably still a library of prompting tutorials, faithfully maintained, teaching everyone to do the thing that once paid $335,000. That was the librarian's actual job description, and the librarian did it well. You know a priesthood succeeded when the scripture works without the priest.
webmaster@ standardization.Operation dissolves. Judgment compounds. The judgment tier needs infrastructure the prompt never did.
"Knows what the output is worth, what it's for, and what it must never do" is the part of this work that survives, and it isn't a phrasing skill. It needs a record of what produced a result, a way to say how much a given agent's output is worth, and a check that the constraints held. The agent trust stack is that layer: provenance, ratings, and verification as installable pieces rather than things each team reinvents once prompting is table stakes.
Read the Theory of Agent Trust
pip install agent-trust-stack · npm install agent-trust-stack
Or the pieces on their own: pip install chain-of-consciousness / npm install chain-of-consciousness for provenance, pip install agent-rating-protocol / npm install agent-rating-protocol for ratings.