The Prompt Engineering Job Was Never Real
It was a hype artifact stretched across an eighteen-month window where models were powerful enough to be useful and dumb enough to need a babysitter. That window closed.
In the summer of 2023, "prompt engineer" was the hottest job title in tech. Anthropic famously advertised a role at up to $335,000 a year. LinkedIn feeds filled with $10 PDF guides. Bootcamps minted certificates. Consultants rebranded overnight.
Three years later, the title has quietly disappeared from most job boards, salary bands have collapsed, and Microsoft's own survey of 31,000 workers found "prompt engineer" ranked second-to-last among roles companies plan to hire in the next eighteen months.
Here is the uncomfortable take: this was not an industry that died. It was an industry that was never born. The title was always a mirage — and the people who benefited most were the ones who noticed it was a mirage the earliest.
The numbers never supported the story
Most of the viral "$300K prompt engineer" headlines traced back to one or two job listings at one or two labs. The dataset never caught up to the hype.
What the job market actually did
Compare that to machine learning engineer, data engineer, or even "AI engineer." Those have thousands of postings, stable salary bands, and a discipline the title actually points at. Prompt engineering never did.
Three reasons the title was always vapor
A real job title points at a stable body of knowledge that takes years to acquire and cannot be absorbed by adjacent roles. Prompt engineering failed all three tests from day one.
What it lacked to ever be a real job
No stable canon
The best practices of Jan 2023 were obsolete by Jan 2024. A discipline you relearn every six months is not a discipline. It is a release note.
No moat versus adjacent roles
Every ML engineer, product manager, and technical writer was already doing prompt engineering. They just called it writing, testing, or spec work.
Models getting smarter ate the job
Each model generation needed less coaxing. A GPT-2 whisperer had real leverage. A Claude 4 whisperer has a productivity tip.
The uncomfortable truth is that prompt engineering was a niche skill in a short window when models were smart enough to be useful but stupid enough that phrasing dominated outcomes. That window is closed for frontier models and closing fast for everyone else.
What a real discipline looks like vs what prompt engineering was
It is worth spelling out what separates a career track from a skill. The difference is not prestige or salary. It is durability.
Real discipline vs. vibes-era job title
| Feature | Real engineering discipline | Prompt engineering circa 2023 |
|---|---|---|
| Body of knowledge | Decades of papers, textbooks, standards | Twitter threads and screenshotted system prompts |
| Years to competence | 3 to 10 | A weekend |
| Tools | Reproducible, versioned, testable | Paste-and-pray in a chat window |
| Transferable across shops | Yes — SQL is SQL, TCP is TCP | No — every model had its own folklore |
| Career ladder | Junior, senior, staff, principal | Posted on LinkedIn, ghosted six months later |
None of that is an insult to the people who were hired as prompt engineers. Most of them were smart, curious, and early to a real shift. The title was the problem, not them.
The skill is real. The job was not.
What is left now, after the title evaporates, is actually more valuable — and more demanding. The people who thrived did not stay prompt engineers. They absorbed adjacent work until the prompt was one small slice of a much larger craft.
What the prompt engineers who survived pivoted into
Context engineering
Designing the information environment an LLM or agent operates in — retrieval, memory, tool schemas, few-shot selection. The prompt is one component.
Evals and measurement
Building the test harnesses, grading rubrics, and regression suites that turn vibes work into shippable software.
Agent architecture
Multi-step workflows with retries, schemas, and tool use. The unit of design is the pipeline, not the paragraph.
Product thinking
Translating fuzzy model capabilities into concrete user flows. The prompt is downstream of the UX decision.
Notice the common thread: each of these requires engineering fundamentals that predate LLMs. Distributed systems. Software testing. API design. UX. The magic of the prompt-engineer era was that for a brief moment those fundamentals looked optional. They were never optional. They were just briefly undervalued.
So what do you do now
If you were studying prompt engineering to land a six-figure remote job: the job is not coming. Pivot. Learn evals, retrieval systems, and how to wire up tool calls. These are the skills the 2026 AI engineer market is actually paying for.
If you hired prompt engineers: re-title them. AI engineer or applied AI gives them a career ladder and you a real talent pool. Stop writing job descriptions that point at a skill instead of a role.
If you are a recruiter screening AI candidates: ask what they evaluated, not what they prompted. Anyone can show you a clever system prompt. Only the real ones can show you the eval set that proves it works.
The prompt engineering job was never real. The people who were hired under that title mostly were. Their value did not go away when the title did — it just moved to the places where the work always actually was.
More essays like this
I write about AI, careers, and the gap between what trends on LinkedIn and what actually ships.
Read more at kennethraj.net