The Prompt Engineering Job Is Already Dead
In 2023, Anthropic posted a job listing for a prompt engineer. The salary: $175K to $335K.
By 2026, Glassdoor pegs the median at $126K. ZipRecruiter's average is $70K. That's not a market correction. That's a skill being absorbed into the bloodstream.
And the irony is almost too perfect. Anthropic, the company that posted that famous $335K role, went on to build Claude Code. A product that succeeds through system integration, tool orchestration, and context management. Not clever prompting. The role they helped define turned out to be the one they proved wasn't needed.
The Job Market Collapsed
Job postings for "prompt engineer" peaked in April 2023, then fell off a cliff. By 2025 they'd plummeted to 20-30 per million job searches. Indeed's assessment was blunt: the category was "never large enough to track mathematically."
Microsoft surveyed 31,000 workers about their hiring plans. "Prompt Engineer" ranked second to last. Second to last.
Meanwhile, half a million students enrolled in prompt engineering courses on Coursera. No globally recognized certification exists for the thing they're studying. I keep thinking about that. Half a million people training for a job title that companies already stopped posting.
The Star Trek Prompt
This is where it gets genuinely strange.
VMware researchers Rick Battle and Teja Gollapudi ran an experiment using DSPy: let AI auto-generate its own prompts, then compare them against carefully human-crafted ones. The machine-generated prompts won by up to 24%.
But the performance gap wasn't the interesting part. The interesting part was what the optimal prompts looked like. Bizarre by any human standard. One used a Star Trek reference to improve math performance. Rick Battle: "I literally could not believe some of the stuff that it generated."
The best prompts aren't the ones a human would write. They're the ones no human would think to write. That should bother anyone who built a career around crafting the perfect instruction, because the machines just casually proved they're better at talking to themselves than we are at talking to them.
Domain Knowledge Wins
Harvard and the Boston Consulting Group ran a study with 758 consultants. Gave everyone access to GPT-4 and measured performance. GPT-4 users completed 12.2% more tasks, 25.1% quicker, at 40% higher quality.
Here's the thing though. The study gave some consultants prompt engineering training and others none. Both groups outperformed the no-AI group, but the gap between trained and untrained prompters was surprisingly small. What actually mattered was domain knowledge. The consultants who understood their subject deeply got more out of the AI than those who just knew the syntax tricks.
Ethan Mollick from Wharton has been saying this for a while. His advice on prompting: "Treat it like a person and you're 90% of the way there."
I think he's right. The obsession with prompt syntax, the brackets, the system messages, the elaborate role-playing setups, all of that was a distraction. Knowing what you're actually trying to accomplish matters more than knowing the right incantation to get there. It always did.
What Took Its Place
Sam Altman said it directly: "I don't think we'll still be doing prompt engineering in five years." Andrej Karpathy endorsed a replacement, "context engineering," because "people's use of 'prompt' tends to (incorrectly) trivialize a rather complex component."
Context engineering. Workflow engineering. AI orchestration. Evaluation engineering. Tool design. These are all basically the same thing underneath: building systems, not writing instructions.
Fast Company compared prompt engineering to being an "Excel wizard." A valuable skill, but not a job title. We all got good at Excel eventually. Nobody kept a dedicated Excel person on staff. (My first real office job had one, actually. Nice guy named David. He got reassigned to operations within a year.)
The models got better at interpreting intent. The tooling got better at structuring context. The gap between a novice prompt and an expert prompt shrank to almost nothing.
The Actual Skill
Prompting was never the thing. It was a temporary stand-in for something harder to name. Understanding the problem deeply enough to know what the machine needed to hear. That's not a technique you can teach in a Coursera course. That's judgment. Domain expertise. Years of actually doing the work.
The $335K prompt engineer and the $70K prompt engineer probably write similar prompts by now.
The difference between them was never the prompting. I don't think it ever was.