Prompt Engineering Was a Job Title for Eighteen Months

In the spring of 2023, a job listing from Anthropic circulated on LinkedIn faster than most people could read it. The company was hiring a Prompt Engineer and Librarian, and the salary band reached comfortably into six figures. Screenshots travelled through group chats with a shared subtext: here was proof that talking to a chatbot had become a profession, and a well paid one. Within a year the same platforms were full of prompt engineering bootcamps, Notion templates sold for the price of a paperback, and consultants who had rebranded overnight. By early 2025 the listings had thinned to almost nothing. The role did not fail so much as evaporate, and the reason it evaporated tells you more about language models than its brief boom ever did.

A skill invented by scarcity

Prompt engineering existed because early models were unreliable narrators of their own capabilities. GPT-3, and the first public builds of tools like Midjourney, would do remarkable things and then, on a near identical request, produce nonsense. The models were powerful but brittle, and the gap between what they could do and what they would do on any given attempt was wide enough to stand in. People who learned to close that gap looked like specialists. They knew that a poorly phrased instruction returned a hedged, useless answer, while the same request framed as a worked example produced something usable.

The craft was real. Few shot prompting, where you show the model two or three examples before the actual task, genuinely lifted output quality. Chain of thought, the discovery that asking a model to reason step by step improved its arithmetic and logic, was a legitimate research finding before it became a LinkedIn slide. For a stretch, knowing these techniques was the difference between a demo that worked and one that embarrassed you in front of a client.

The tricks that stopped working

The trouble with folk techniques is that they date quickly, and this folk knowledge dated within months. A generation of prompts told the model it was a world class expert, promised it a tip, or warned it that a colleague would be fired if the answer was wrong. These theatrics produced measurable gains on weaker systems. On GPT-4, on Claude, on Gemini, the same incantations did close to nothing, because the models had been trained on enough human feedback to infer intent from a plain request. The scaffolding you built to compensate for a model's weaknesses becomes dead weight the moment the weakness is patched.

This is the pattern that undercut the whole discipline. Every clever workaround was a bet that a specific model limitation would persist. Model providers were, at the same time, spending enormous effort to remove exactly those limitations. Instruction tuning, better refusal handling, longer and more coherent context windows: each release quietly retired a chapter of the prompt engineering handbook. The people who had memorised the tricks were, in effect, shorting their own expertise.

Absorbed into the job everyone already had

What replaced the standalone role was not nothing. It was the realisation that the useful part of prompting was never the magic words. It was knowing precisely what you wanted, describing it without ambiguity, and checking whether you got it. Those are the skills of a competent product manager, a careful technical writer, or an engineer who writes good tests. They did not need a new job title because they already had several.

Inside teams that ship with language models, the conversation moved on. System prompts became configuration files, versioned in Git and reviewed like any other code. The interesting work migrated to evaluation: building datasets of real inputs, scoring model outputs against them, and catching regressions when you swap one model for another. Writing a clever prompt is an afternoon. Building an eval suite that tells you whether your product still works after a model update is a discipline, and it is one that sits with engineering, not with a lone prompt whisperer.

What actually survived

Strip away the hype and a durable competence remains. It is closer to interface design than to incantation. You are specifying a contract with a system that reads natural language, and the quality of that specification still matters enormously for agents, for retrieval pipelines, and for anything where a model calls tools on your behalf. But this is context engineering, spec writing, and evaluation, distributed across the people who were already doing adjacent work. It is a layer of every technical job now, not a job of its own.

The eighteen month lifespan of prompt engineering as a title should be read as a small, clean case study in how quickly automation reshapes labour when the tool itself is the thing improving. The role appeared to bridge a gap, the gap narrowed under it, and the bridge was quietly dismantled. Expect the same arc to repeat wherever a job forms around coaxing better behaviour out of a young technology. As models grow more capable of inferring what we mean, the premium shifts from knowing how to ask toward knowing what is worth asking for, and that is a harder skill to package, sell, or make redundant.