What was said
In 2023, "prompt engineer" was seriously treated as an emerging profession: companies posted six-figure salaries, courses and certifications bloomed, and media called it the ticket into the AI industry for people who don't code. The argument: the model's capability is locked inside the prompt, so writing prompts well is using AI well, and the skill can only appreciate.
What actually happened
Within two or three years, the standalone occupation had essentially evaporated. The reason is direct: models learned to understand plain speech. Reasoning models decompose vague requests themselves, ask clarifying questions, and fill in context — the incantations that once needed careful construction now work in everyday language. What little prompt craft remains became a default literacy for every job, the way "can use a search engine" was never an occupation. The few roles that did survive renamed themselves (AI product, evaluation, agent engineering) and do entirely different work.
Why it didn't come true
They mistook a transitional friction for a permanent skill. Every technology transition has this window: early electric factories needed specialists to tend the belt drives; early cars needed a driver-mechanic. The wage premium inside the window is real, but its substance is rent on a bad interface. When the interface improves, the rent goes to zero. The law is much older than AI; each generation just mistakes its own window for the end state.
The unexpected part
What actually appreciated was not "can talk to a model" but "knows what to ask for, and can judge whether what came back is right" — taste and acceptance ability. When everyone can make the model produce things, the scarce resources became selection, standard-setting and answering for outcomes. Nobody wrote that into the predictions at the time, because it doesn't sound like an "AI job."