Screening
Prompt Engineering Survived as a Skill and Died as a Job Title
Prompt engineering is still a real skill, and it also stopped being a job of its own; the two answers are not in conflict. The work moved inside ordinary roles, so it belongs in a job description as something a person does with your material. Its content moved too: models got better at inferring intent, and the craft shifted from clever phrasing to supplying the right context and checking the output. Screen it by watching somebody brief a model on a task from your own work.
The takeThe essays declaring prompt engineering dead and the listicles quoting prompt-engineer salaries are wrong in the same direction, because both are arguing about a title while the skill quietly became a condition of employment in most desk jobs. Screening on the title selects for whoever updated their headline first. That is real information about a person, just not the information being hired for. The argument over whether the title lived or died was never a hiring question, and the people who won it are no better at telling one candidate from another.
Where Olive fits
Open a role and see what the work shows
Screening on a posting keyword tells you which phrase a candidate adopted. Olive assesses the person instead: a role-grounded assignment done with an AI assistant, returned as six findings a human wrote, each anchored to the moment in the session it came from.
Rank your shortlistIs prompt engineering still a real skill?
Still a skill, yes, but now a component of other jobs rather than a job of its own. The specification part of working with a model, meaning what context goes in, what the output has to satisfy and what counts as done, still separates a useful first pass from a long correction loop. What went away was the standalone role, along with the idea that phrasing tricks were the substance of the thing.
The vocabulary spread much faster than the roles did, which is what makes a title screen misleading. Indeed Hiring Lab counted job titles carrying an AI term in the employer's own raw title: US AI-touched titles rose from 264 in Q1 2022 to 822 in Q1 2026, 8.3% of all titles with enough postings to measure, and 63% of them now sit outside tech occupations 1. That counts distinct titles rather than postings or hires, the inclusion threshold is five postings in a quarter, and the count dipped in early 2023 before climbing, so a straight line through the two endpoints would be wrong.
Employers are pricing the skill into what they advertise at the same time. Lightcast reports that salaries in postings mentioning AI skills run 28% higher than postings that do not, roughly $18,000 more a year, and that in 2024 over half of postings requesting AI skills sat outside IT and computer science 2. That is a raw comparison between two groups of postings rather than a like-for-like premium, AI-mentioning postings skew senior and toward higher-paying industries, and it measures advertised pay rather than anyone's actual salary. It still says the language is everywhere. Language being everywhere is exactly why matching on it screens for adoption of a phrase.
What changed inside the skill?
The center of gravity moved from phrasing to context and checking. Early prompt craft was largely about coaxing a weaker model: role-play preambles, formatting incantations, instructions to think step by step. Models got better at inferring intent, which retired most of that. What did not retire is supplying the constraints and the source material, and reading what comes back with suspicion.
A field experiment run with 758 Boston Consulting Group consultants put the point sharply. On the task deliberately placed outside the model's capability, consultants using GPT-4 were 19 percentage points less likely to reach the correct answer, and the subgroup given a prompt-engineering overview did worse than the subgroup given none 3. One task, one sample, a 2023 model, so this is not a general finding about coaching. The direction is still worth sitting with: instruction on how to ask did not help on the task where asking was the wrong move, because the missing competency was judgment about the output rather than craft in the request.
So the modern content of the skill looks less like an incantation and more like briefing a contractor. What is this for, who reads it, what is already known, what would make the answer wrong, and what am I going to check before this leaves my desk. The last two are where it stops being a phrasing skill at all, which is the competency worth weighting in a hiring bar.
Write the requirement as a behavior, not a title
Give the requisition one line, phrased as something a person does with your material. A requirement reading "prompt engineering experience" selects for whoever adopted the phrase. A requirement reading "briefs a model with the context a colleague would need, and checks the output against a source" selects for the work. The second version is also the one an interviewer can grade consistently against a rubric.
| Instead of | Write |
|---|---|
| Prompt engineering experience required | Briefs a model with the context a colleague would need to do the task |
| Three or more years working with LLMs | Has done this job with an assistant available, and can say what they kept for themselves |
| Familiar with the major AI assistants | Names the last thing an assistant got wrong in this domain, and what they checked it against |
| Prompt engineering certificate preferred | Delete the line, and run the exercise in the next section instead |
The years-of-experience line deserves its own funeral. There are only a few years available to have, and the tools of two years ago barely resemble the current ones, so the filter selects on when somebody started rather than on what they can do. A certificate line has the same problem with an extra one attached: it imports somebody else's definition of the bar, usually built around a single platform. Whether a prompt engineering certificate is worth anything next to a work sample works through the trade honestly.
What the requirement should reflect is that skills are changing rather than vanishing. Indeed's skill-transformation index rated almost 2,900 work skills and mapped them onto more than 53.5 million postings: 40% fall into minimal transformation, 19% assisted, 40% hybrid and 1% full 4. Those ratings come from language models scoring skills rather than from anyone observing work, so they describe what models are judged capable of, and hybrid explicitly means human oversight is still required. Hybrid is what the specify-and-check pattern looks like in a posting.
How do you screen for it in twenty minutes?
Hand over a task from the role and read the brief rather than the output. Give the candidate a real request, the source material and an assistant, and ask them to reach a first draft. What carries the signal is the first three turns: whether the constraints went in before the ask, and whether anything came back that they declined to use.
1. Pick a task the role does weekly. Real material, real constraints, and one thing in the packet that is confidently wrong in a way this field would catch. 2. Twenty minutes, assistant available, and say so in the invitation. A candidate who thinks the tool is banned will perform compliance instead of work, and you will learn nothing. 3. Read the first three turns for framing. Did the constraint, the audience and the source material go in before the request for output, or did the ask arrive first and the context dribble in over the correction loop. 4. Read what got cut. Anything the model produced that did not make the draft, and the reason. This is the part that separates two candidates whose drafts look identical.
What not to grade: prompt syntax, the number of turns, which assistant they picked, and how fast the draft arrived. None of those survived contact with better models, and the last one rewards skipping the check. A tool list on a resume is a starting point for a question rather than evidence, which is the whole of how to tell whether a candidate really uses the tools they list.
Common questions
Should I hire a dedicated prompt engineer?
Rarely, and the exceptions are narrow. A team shipping a product whose behavior depends on prompt design, evaluation sets and regression testing has a genuine specialist job there, and it is closer to applied evaluation engineering than to writing clever requests. For everything else the skill belongs inside the role that owns the output. A standalone title tends to produce one person who owns the phrasing and nobody who owns whether the answer was right.
Is a prompt engineering certificate worth anything?
As evidence that a course was completed, yes. As evidence of judgment, no. It records that somebody sat through the material, which leaves open the question you actually care about, whether this person has ever caught a wrong answer in your domain. Treat it the way you treat any completion certificate: a conversation starter that costs zero minutes of grading, and a poor screening criterion, since it filters on who could pay for a course.
Should prompt engineering appear in the job title?
Not unless the role genuinely is that. A title carries into the applicant tracking system, the org chart, the internal salary band and the person's next job search, and it is expensive to change later. If the work is a marketing analyst who uses an assistant daily, the title is marketing analyst and the AI part belongs in the responsibilities. Titles that chase a trend age badly and make internal mobility harder for the person who holds them.
What about a years-of-AI-experience requirement?
Drop it. The assistants of three years ago behave so little like the current ones that time served with them measures almost nothing, and the line quietly filters out people who came to the work from a different field last year, which is most of the strongest candidates in non-technical functions. If depth is the worry, a twenty-minute exercise on real material from the role measures it directly.
Is prompt engineering the same thing as AI fluency?
No. It maps roughly onto one of the four competencies in the AI fluency framework, the one about describing a task well enough to get a useful result. The other three, deciding what to hand over in the first place, judging what came back, and vouching for what ships, are not phrasing skills at all. Using the two phrases interchangeably in a posting is how a requirement ends up testing a quarter of what the role needs.
References
- 1. AI Is No Longer Just a Tech Occupation Story: It's Spreading Across Job Titles in the US and Europe hiringlab.indeed.com Supports the claim that AI vocabulary spread across job titles and outside tech: 264 titles in Q1 2022 to 822 in Q1 2026, 63% outside tech occupations.
- 2. Beyond the Buzz: Developing the AI Skills Employers Actually Need lightcast.io Supports the claim that employers already price AI skills into advertised pay, and that most postings requesting them sit outside IT and computer science.
- 3. Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality (Working Paper 24-013) mitsloan.mit.edu Supports the claim that prompt-engineering coaching did not help on the task outside the model's capability, where the subgroup given it did worse than the subgroup given none.
- 4. AI at Work Report 2025: How GenAI is Rewiring the DNA of Jobs hiringlab.indeed.com Supports the claim that most skills are changing rather than disappearing, with hybrid transformation explicitly meaning human oversight is still required.
4 sources, numbered by first appearance. How Olive sources claims
General guidance for hiring teams. What works at one company and one volume may not transfer to yours.
Olive assesses how a person works with AI. It does not detect AI-written documents, and it never produces a score, a ranking, or a match percentage for a person. Candidates read the same report the employer reads.