Prompt engineer: what the job is, what it pays, and whether it still exists

In short: a prompt engineer designs and debugs requests to AI models so they reliably solve a real task. As a standalone job title the role has largely dissolved — the skill moved inside other roles. But there's something more interesting than that. On 17 July 2026 we audited the market of courses selling this "profession", and found that courses with the words "prompt engineering" in the title teach almost anything except how to debug a prompt. Below: our numbers, the verbatim reviews, and what to learn instead.

What a prompt engineer actually does

The title misleads. It sounds like someone spends the day composing elegant phrases for a chatbot. In practice the work is closer to engineering debugging. The brief is: "make the model produce a usable result across a thousand different inputs, not just on your one lucky example." Hence the actual work:

• Framing the task. Turning a vague "let AI handle customer requests" into a testable requirement with clear good/bad criteria.
• Designing the prompt. Role, context, format, examples, constraints — the basics from what a prompt is and how to write one.
• Testing. Dozens of real cases, a run, a comparison between prompt versions. Without this you're guessing.
• Fighting failures. The model invents facts, ignores the format, breaks on long inputs, drifts from the instruction. All of it gets fixed in the prompt, the data, or the architecture.
• Wiring in data and systems. Often the answer must rest on the company's knowledge base, not the model's memory.

Think before reading on: if the model got your task wrong in one case out of twenty — how would you even find out which one? Being able to answer that is half the profession. And as we're about to show, it's the one question almost no course on the market answers.

Where the hype came from and why it deflated

In 2023 the models were more temperamental and the market was disoriented. A couple of loud job ads with six-figure dollar salaries made the headlines, and a sense of a new gold rush appeared: "type text into a chat, get paid like a senior engineer." Then three things happened.

1. Models got smarter. Current versions understand plain human phrasing far better. Magic incantations like "you're a genius and my career depends on this" stopped producing any visible effect.
2. The skill turned out to be literacy, not a profession. Like knowing how to use search or a spreadsheet: critically useful, but nobody hires a "search operator" as dedicated headcount.
3. Real problems turned out to be bigger than the prompt. A business doesn't need request text; it needs a working system with data, quality evaluation, cost control and safety. That's a different role.

The hype deflated in hiring. It did not deflate in the business of selling training. That part is alive and well — and it's the only part of the story you can actually measure. So we measured it.

Our research: how many people have already "trained as a prompt engineer"

On 17 July 2026 we pulled data from Udemy (through the internal API behind their course pages) and from Coursera's review pages. Here are six courses devoted entirely to prompting. Four carry the words prompt engineering right in the title; the other two — "Google Prompting Essentials" and "Prompt and Context Engineering 101" — put prompting in the title using different words. We'll count both ways.

• Prompt Engineering for ChatGPT (Vanderbilt, Jules White) — 698,444 enrolled, 7,915 reviews, rated 4.8, 19 hours.
• Generative AI: Prompt Engineering Basics (IBM) — 654,256 enrolled, 2,313 reviews, 4.7, around 10 hours.
• Google Prompting Essentials (Google) — 365,425 enrolled, 7,219 reviews, 4.8, under 10 hours.
• Prompt Engineering Specialization (Vanderbilt) — 138,967 enrolled, 9,151 reviews, 4.8, 39 hours.
• Prompt Engineering with ChatGPT Masterclass (RPATech, Udemy) — 118,803 students, 57,836 reviews, 4.48, 4 hours, 41 lectures.
• Prompt and Context Engineering 101 (Mike Wheeler, Udemy) — 84,942 students, 24,401 reviews, 4.31 — the worst rating in our sample.

Add them up: those six come to 2,060,837 enrolments. Be strict and keep only the four with prompt engineering literally in the title, and you still have 1,610,470. The conclusion doesn't shift an inch either way. For scale, the biggest non-technical course on Coursera, Google AI Essentials, has 1,876,929 enrolled; Andrew Ng's Generative AI for Everyone has 814,083. On Udemy the leader, Generative AI for Beginners by Aakriti E-Learning, has 409,492 students against 121,079 reviews and a 4.53 rating.

This matters more than any salary range. If "prompt engineer" were a scarce profession, two million graduates of courses about exactly that skill would have filled it by now. The niche isn't empty — it's crowded. Walking in with "I completed a prompt engineering course" means joining a queue with hundreds of thousands of people holding the identical line.

A caveat against ourselves: "enrolled" on Coursera and "finished" are not the same thing, and neither platform publishes completion data. So two million is an upper bound on competition, not a count of practitioners. But the lower bound is clearly not "shortage" either.

The prerequisites: no code required, anywhere

We read the requirements block on the top courses. Verbatim: "No prerequisites as ChatGPT is a tool anyone can access and use immediately" (Ballinger). "No coding skills needed" (Wheeler). "A desire to learn" (Aakriti). "No prior experience with AI or programming is needed" (Complete AI Guide).

Read that as a market signal, not an invitation. An entry barrier equal to "a desire to learn" means the course filters nobody — and therefore its certificate carries no information for an employer. A profession you can enter with zero prerequisites doesn't get paid like a profession. That isn't a moral judgement; it's supply-side arithmetic.

It also tells you what the courses are really selling. Not a job. Access to a tool that already has no access barrier.

What it costs: there is no Udemy discount

A separate finding, and the reason we went into the API in the first place. Everyone knows about Udemy's permanent sale — a marketplace storefront is very good at presenting a price as a vanishing opportunity. Rather than argue with the storefront, we requested the raw response from their own pricing endpoint. Here it is, unprocessed:

• "price": {"amount": 24.99, "currency": "EUR"}
• "list_price": {"amount": 24.99, "currency": "EUR"}
• "saving_price": {"amount": 0.0}
• "has_discount_saving": false
• "discount_percent": 0

The price equals the list price. Saving: zero. Discount percent: zero. The has-a-discount flag is literally false. Whatever the countdown timer on the page is doing, it has no relationship to these fields.

The real figures as of 17.07.2026: 11 of the 12 top courses cost €19.99; one (Complete AI Guide) costs €24.99. Udemy Personal Plan is €20.00/month, €10.00/month on promo. Coursera Plus is €50/month or €343/year with a 14-day refund window. Google's own programmes on Coursera run $49/month after a 7-day trial.

The practical takeaway: don't let the timer rush you. You pay €19.99 today, €19.99 tomorrow and €19.99 next month. The only thing the countdown accomplishes is stopping you from calmly reading the reviews. And the reviews are worth reading.

The core finding: courses with "prompt engineering" in the title don't teach prompting

We went through the reviews and grouped the complaints. The densest cluster is people who bought a course for exactly the skill in its title and didn't get it. Verbatim, with authors and star ratings:

• "There is nothing teached about creating a good prompt. It is just an overview of types of prompts" — Geralt O., 01.07.2026, 2★ (Wheeler).
• "No specific guidance on prompt engineering… what to avoid while asking, how to organize your thoughts, how to give feedback to AI based on its answers" — Bharat Ram A., 05.06.2026, 1.5★ (Wheeler).
• "i thought it would go deeper in prompts and have more examples and sessions to master or enhance our current prompts" — Manuel L., 15.06.2026, 2★ (PE for Everyone).
• "Too much background on ChatGPT. Get me to how to prompt GPT" — Mark R., 16.09.2025, 1.5★ (Academind).

Look closely at Bharat Ram A. He lists three things: what to avoid when asking, how to organise your thoughts, and how to give feedback to the AI based on its answers. That third item is debugging. A paying student described the hole in the market in his own words, because he went looking for it and it wasn't there.

Manuel L. wants to "master or enhance our current prompts" — same request, different words. Enhance an existing prompt. Not learn a formula: fix a thing that already exists and doesn't work well enough. Four independent people, four different courses, one missing lesson.

And here's the complaint that explains how it got this way. About the most-enrolled prompt engineering course on Coursera:

"I finished week 2. And all the lessons could be in 1 lesson. I jumped to week 3, 4, 5 and 6, and I see some concept reapeted" — Shai Mizrachi, 22.07.2023, 1★ (Vanderbilt).

And on the delivery format in general:

"why read straight from the slide? I can do that. This was not a helpful course at all" — Janie I., 02.07.2026, 1★ (Justin Barnett).

Worth pausing here. "I can do that" isn't a complaint about a bad lecturer. It's a statement that video which is merely a spoken slide loses to text by definition: text comes with the reader's own pace and a transcript for free. A student paying 42 hours of life to the Complete AI Guide is buying, in time, something they could have read in one.

Two more, same cluster, different courses: "50% of this course is reading script like a robot from the slides" — Shashank T., 13.04.2026, 1★. And on the RPATech course with 118,803 students: "Written by AI, delivered by AI. The slides are way too crowded to be useful" — Bradley M., 16.04.2026, 1★.

The structural hole: nobody checks that you learned it

Now the part that matters most for this topic. A prompt engineering course is obliged to check prompts. There is exactly one way to check a prompt: run it on a model and look at the answer. A video platform has no model inside the lesson. So there is nothing to check with. Here's what comes out the other end:

"you send your assignments and immediatly you got your results: 100% correct. I am still speechless. I put all that effort in and have no idea whether I my answer was correct or not" — Shinysheep, 21.09.2023, 1★ (Vanderbilt Prompt Engineering).

That is not a lazy instructor. It's a consequence of architecture: the platform physically cannot evaluate a prompt, so it awards a pass. The neighbouring complaints all rhyme:

• "Videos are too short and superficial so you end up memorizing sentence by sentence to pass quizzes. Not a learning experience" — Laurie J Phillips, 12.04.2024, 3★ (IBM).
• "quizzes give unhelpful feedback for incorrect answers and just say 'watch the video again'" — Cory Covino, 04.05.2024, 2★ (IBM).
• "All the coding is done in the labs for you. You won't have to debug anything or figure anything out, just press shift-enter" — Cornelius Griggs, 1★ (Generative AI with LLMs).

Griggs is describing the same disease in the lab format: the exercise runs, so it feels like practice, but nothing was ever at stake. "Just press shift-enter" and "100% correct" are the same event.

And the kicker. Vanderbilt's Prompt Engineering Specialization requires a paid ChatGPT+ subscription to complete the assignments. A course you're already paying for through Coursera Plus (€50/month) sends you outside for the model. Student economics: roughly €70/month — to watch videos and receive an automatic "100% correct." On the exact price of the ChatGPT subscription we'll be straight with you: we couldn't get it from the primary source, every OpenAI domain returned 403, and the ~$20 figure is secondary. The €50 for Coursera Plus is off their own page.

The content rots, and students notice

"Most content is from 2024. This course is not bad for its time, but just too dated now" — Martin F., 27.05.2026, 2★ — on a course whose card says "updated 04.2026."

"The content is mostly from 2023. …I invested my 41 hours and Im learning content which is from 2023" — Harsh A., 09.04.2026, 1.5★ (Complete AI Guide).

Sometimes stale simply becomes wrong: "Stable diffusion, Dalle and Midjourney are not GAN architecture powered. They're diffusion models" — Jose C., 12.05.2026, 1★. The student is correcting the course, and the error lives not only in the lecture but in the quizzes — meaning a correct answer is marked wrong.

The cause is boring and structural: re-shooting video is expensive, editing text costs almost nothing. In a field where the tools change every quarter, that difference decides everything. Note the gap between "updated 04.2026" on the card and "most content is from 2024" in the review — a card date measures when someone touched the course, not when the content was true.

About salaries — honestly

You've surely seen articles with specific figures. Treat them carefully: the numbers usually come from a handful of loud postings at the peak of the hype, from aggregators with tiny sample sizes, or from reports by companies that benefit from interest in the topic. There is no reliable average "prompt engineer salary" today — because there is no stable, widespread job title by that name. We are deliberately publishing no salary number here: we have no source we'd stand behind, and inventing a plausible one is exactly what the articles we're warning you about do.

What can be said without lying: you get paid for a role, not for prompts. ML engineer, data scientist, AI developer, product manager with an AI specialisation — those positions have a real market and real ranges, and prompt skill is one requirement among many, not a substitute for the rest. If you see "become a prompt engineer in a month and earn X" — that's selling a myth, not career advice.

You can check this in five minutes without taking our word for it: open any job board and search the exact phrase "prompt engineer", then "AI engineer" or "machine learning engineer". The difference in result counts tells you more about the market than any salary survey. While you're there, read the requirements on the few postings that do turn up: almost certainly they'll list code, data and domain experience, with prompting itself as one bullet out of ten.

Put that next to our enrolment numbers and the picture closes. Supply: 2,060,837 enrolments on courses teaching exactly this skill. Demand: a search term that barely returns results. Nothing about that combination produces a salary.

What to learn instead of the "profession": debugging

Let's pull the findings together. Everyone teaches "here's the prompt formula." Almost nobody teaches: the prompt didn't work — what do you change first? The four reviews above are asking for precisely that. Which makes it the skill that separates you from two million certificate holders.

Prompt debugging is a loop, not a flash of insight:

1. Name the symptom precisely. Not "bad answer" but: invented a fact; ignored the format; too generic; wrong tone; broke on a long input; answered a different question. Different diseases, different cures. Most people stop at "bad" and then rewrite at random.
2. Change one thing at a time. Edit five things and you'll never know which one worked. It's literally the same rule as in any debugging.
3. Test on several inputs. One good answer isn't a result, it's a coincidence. Models are non-deterministic; a single run is a sample of size one.
4. Define what "good" means. Until there's a criterion, you're not debugging — you're reshuffling wording until you like it. This is the step that separates the job from the hobby.
5. Version your prompts. v1, v2, v3, and what exactly changed between them. In a week you won't remember.
6. Know when the prompt isn't the problem. Missing data is not a wording issue. No prompt retrieves a fact the model was never given. Recognising this early saves days.

And now the thing Udemy and Coursera physically cannot offer. Below are working prompts. Every block has a Run button: press it right here and watch a live answer come back. That is the check a video platform cannot install — not because it doesn't want to, but because there's no model in the lesson.

Drill 1: dissecting a bad prompt

A genuinely weak prompt is already pasted in — the kind everyone writes at the start. Run it and see what comes back.

You are a prompt debugging instructor. Dissect the prompt below.

PROMPT UNDER REVIEW:
"Write a post about our new product. Make it look nice and professional."

Do exactly four things:
1. List the information that is missing for any answer to be good at all.
2. Name the criteria stated here that cannot be verified, and explain why.
3. Predict three specific ways the answer to this prompt will be bad.
4. Give an improved version plus a table: what I changed -> why.

Do not compliment me. Be concise.

Drill 2: predicting failures before you run

The move that marks an engineer: think about where it breaks first, run second. The prompt is self-contained — the task is described inside it.

Context: I am about to give an AI this instruction for handling support email.

INSTRUCTION:
"Read the customer's email and classify it as: refund, delivery,
technical issue, or other. Answer with one word."

Do not follow that instruction. Run a risk review instead:
- Five ways it will produce a wrong result on real email.
- For each, an example email (2-3 sentences) that triggers it.
- Exactly what to add to the instruction to close each case.
- Separately: which of these problems CANNOT be fixed by prompting, and why.

Format: a table.

Drill 3: a test set and a definition of "good"

The step the courses skip and the employer checks.

Help me build a test set for a prompt.

The prompt's job: turn raw meeting notes into a list of action items
with an owner and a deadline.

Give me:
1. Ten test inputs - short meeting notes, 3-5 lines each.
   Make three of them edge cases: no deadlines mentioned; a contradiction;
   a discussion where nothing was actually decided.
2. For each input, what counts as a correct answer.
3. A rubric of five criteria scored 0/1 each, with a pass threshold.
4. One input where the correct behaviour is to refuse and ask a question.

Write it as a finished document, no preamble.

Drill 4: a live exam on the skill

Test yourself. There will be no automatic "100% correct" here.

Examine me on prompt debugging. Rules:

- You give me ONE case: a task, a prompt, and the model's bad output.
- You invent the case yourself: realistic, from office work.
- I answer with what I would change first and why.
- You grade me strictly: what I missed, what I diagnosed wrongly,
  which hypothesis would have been stronger than mine and why.
- Do not agree out of politeness. If I am wrong, say so plainly.
- After grading, give me the next case, harder.

Start with the first case. Do not explain theory up front.

Common mistakes of people "learning prompts"

• Collecting incantations. A list of 500 prompts is not a skill. The skill is fixing the one prompt you actually need. Treat our prompt examples as starting material, not recipes.
• Studying without a task. A prompt with no real problem behind it is an exercise in a vacuum. That's where "only theory is discussed which we already know" comes from — Aditya Nagavolu, 19.11.2023, 1★.
• One run equals a conclusion. The model is non-deterministic. Liking the first answer proves nothing.
• Editing five things at once. A classic. It got better — nobody knows why.
• Believing in the certificate. See the numbers above: 698,444 people on one course. A certificate is not a signal.
• Rushing because of a timer. has_discount_saving: false. The countdown isn't about the price, it's about your attention.
• Confusing length with depth. 42 hours and 545 lectures in one course, and the top complaint is padding: "The course is far too drawn out… padding the course with filler" — Victoria X., 13.02.2026, 2★.

What good prompt people actually become

• AI engineer / AI developer — builds products on top of models: integrations, agents, knowledge-base search. Code required.
• Product or project manager with an AI focus — knows where AI fits, what it breaks and how to measure the benefit. Less code, more product thinking.
• AI quality specialist — assembles test sets, evaluates outputs, catches regressions. Little glamour, lots of value. Note this is exactly the skill the courses replace with an automatic "100% correct" — which is why it stays scarce.
• An amplified specialist in their own field — a marketer, lawyer, analyst or teacher who does their job noticeably faster. The most common and most underrated path.

A separate trend is AI agents: systems that carry out multi-step tasks with tools. There the prompt is only one layer — but nothing works without it. A neighbouring story is vibe coding, where code gets written through dialogue with a model.

How to pick a course, if you still want a course

We read hundreds of reviews and boiled the result down to four questions. Ask them of any course — including ours.

1. Where will I run the prompt? If the answer is "in another tab, on a separate subscription", the course will not check your work. See Vanderbilt and ChatGPT+.
2. Who grades my prompt? If it's a keyword autograder, you'll get "100% correct" and zero information. Shinysheep put "all that effort" in and learned nothing about their own answer.
3. Does it teach debugging? Open the syllabus and find the lesson where a prompt fails. If there isn't one, the course is about a formula, not about the work.
4. When was the content actually updated? Not by the date on the card, but by reviews from the last month. Martin F. found 2024 material on a course claiming an April 2026 update.

And an honest caveat about our own data: the share of negative reviews (≤3.5★) on Udemy is 9.61% for Generative AI for Beginners and 10.36% for the Complete AI Guide; on Coursera it's 1.5–3.5%. In other words the overwhelming majority of students are satisfied, and we deliberately read the minority. We do that not to prove the courses are bad, but because the unhappy minority states most precisely what the product lacks. A happy student says "great course." Bharat Ram A. tells you which lesson was missing. One more limitation worth naming: Coursera's star filter is client-side, so the quotes we could reach come from the default page any visitor sees — we didn't cherry-pick from deep inside a 1★ filter, but we also can't claim we saw everything.

Where to start

1. Take your own task, not a textbook one. The one you do by hand every week.
2. Learn the basics. Role-task-context-format, examples inside the prompt, constraints, permission to say "I don't know."
3. Build a test set. Ten real inputs and a clear idea of what a correct answer looks like. That's what separates an engineer from an enthusiast. Drill 3 above assembles one in a single run.
4. Learn to catch invention. Understanding why AI hallucinations happen, and the techniques against them, is half the practical value.
5. Compare models. The same task behaves differently across them — see the comparison of ChatGPT, Claude and Gemini.
6. Keep a log. The prompt, what failed, what you changed, what came out. In a month that log is your portfolio — and far more convincing than a certificate 698,444 other people also hold.

The takeaway, without hype

"Prompt engineer" as a standalone profession was largely a one-season phenomenon. The skill hasn't gone anywhere — it stopped being a job title and became part of normal work with AI. Our measurements add the second half of the story: the places that teach this "profession" sell the prompt formula to millions of people and don't sell the one thing worth money — the ability to fix a prompt that didn't work. The reason isn't lazy authors. It's that you cannot check a prompt without a live model inside the lesson, and a video platform doesn't have one. So learn where you can press Run. The practical side is in our article on AI for work, the "will I be replaced" anxiety is covered in will AI replace my job, and the fundamentals are in what a prompt is. If you want it systematically, there's our prompt engineering course.