Prompt engineer in 2026: what the job involves, which skills to build and how to show them

Prompt engineer in 2026: a job title that turned into a skill

People who search for prompt engineer usually want two answers: what this person does all day and whether anyone still hires for it. Wikipedia defines prompt engineering as the process of structuring natural language inputs, called prompts, to produce specified outputs from a generative AI model. After ChatGPT was released in November 2022, the article notes, the practice was soon seen as an important business skill, and companies began hiring dedicated staff for it.

The title did not keep its shine. Wikipedia writes that it has since become less common, as models themselves produce better prompts than people, and it cites a 2025 Wall Street Journal piece that called the job one of the hottest of 2023 and now obsolete. The work did not vanish with the label: someone still decides what a model should do in a product, checks that it does so reliably and stops it from doing what it must not.

Four parts of the work hidden behind the title

The first part is writing the instructions: the system message of a support chatbot, the template behind a product description generator, the rules of an assistant that sorts mail. OpenAI’s prompt engineering guide describes the task as writing effective instructions so that a model consistently produces the content you need, and calls it a mix of art and science. One good answer in a chat window proves little.

The second part is testing, because the same guide warns that what a model generates is non-deterministic: one request can return different text on different runs. So the work includes evaluation sets, collections of typical and awkward inputs with a written way to grade the answers. The third part is guardrails, the rules and checks that keep the model inside its task when a user or a document tries to push it out.

The fourth part is connecting the model to the business: its catalogue, its documents, its booking system. Here prompts meet code, data and access rights, and the person in the role works next to developers rather than instead of them. In a small company all four parts often sit with one employee who also does something else.

How OpenAI, Anthropic and Google describe prompting in their docs

OpenAI’s guide recommends pinning production applications to a specific model snapshot so that behaviour stays consistent, and building tests and evaluation suites that measure prompt behaviour, so performance can be tracked as the prompt changes or the model is upgraded. It also separates developer messages, which take priority, from user messages, which rank behind them.

Anthropic’s overview assumes three things before any prompt work begins: a clear definition of success criteria for the use case, some way to test against them empirically, and a first draft prompt to improve. Readers without a draft are pointed to a prompt generator in the Claude Cookbook. The page also says that not every failing test is best solved by prompting: latency and cost sometimes improve more easily with a different model.

Google’s Gemini documentation calls the practice prompt design, the process of creating requests that elicit accurate, high quality responses, and adds a short note that prompt engineering is iterative: its guidelines and templates are starting points, to be refined against the responses you observe. A prompt stays a draft until a test shows that it works.

Evaluation sets: the part of prompt work nobody sees

Anthropic’s page on success criteria and evaluations gives the clearest picture of this work. Good criteria are specific, measurable, achievable and relevant: not good performance in general, but, for example, accurate sentiment classification against labelled examples. Typical dimensions include task fidelity, consistency, tone, privacy preservation, latency and price, and most applications need several at once.

It then sets three rules for designing tests. They should mirror the real task and include edge cases: irrelevant or nonexistent input, input that is far too long, harmful requests and ambiguous cases where people would also disagree. Grading should be automated where possible. And volume beats polish: many questions graded automatically give more signal than a few graded by hand. When one model grades another, Anthropic advises using a different model from the one that wrote the answer.

For an online shop, a support bot needs fifty real customer questions, the same questions rephrased with typos, a few that have nothing to do with the shop and a few that try to extract a discount the shop never offered. Each answer gets a pass or a fail against a written rule, and the set runs again after every change. That table, more than the prompt, shows whether the job was done.

Guardrails, documents and tools around the model

A model connected to nothing can only talk. OpenAI’s function calling guide explains how it reaches external systems and data beyond its training: the application lists the tools the model may use, the model asks for one, the application runs its own code with that input and sends back the result, and the model answers or requests another tool. The prompting guide adds that supplying relevant context with a request is sometimes called retrieval-augmented generation.

Every connection is also a door. Anthropic’s page on jailbreaks and prompt injection separates two threats: a user who deliberately writes inputs to get around the rules, and third-party content, such as a web page, an email, a document or a tool result, that carries hidden instructions while the user is innocent. Its advice includes telling the model that such content is data, not commands, and giving it only the access it needs, so a successful attack can do little damage.

The same page recommends testing your own agent before launch with documents and emails that deliberately contain injection attempts, and monitoring outputs once it is live. Guardrails, in other words, are not one polite sentence in the system prompt but a mix of wording, access rights, checks in code and tests built to break them.

A separate profession or a skill inside other jobs

The Russian job market offers a quick reality check. On hh.ru, a search for the Russian phrase meaning prompt engineer returned 181 vacancies across Russia on 17 October 2026. Among the first twenty results, one was titled prompt engineer. The rest were AI engineers and developers, testers, a systems analyst, machine learning engineers, a project manager, a narrative designer, a neural network operator and an AI intern.

A search page is a snapshot and the count changes daily, but the pattern matches what Wikipedia says about the title. In these results, prompt work shows up mostly inside other roles: analysis, testing, development, management. A dedicated position is the exception. For most readers, then, the realistic path is a stronger version of their current job.

That changes what is worth learning. A marketer who can write a brand voice prompt, run it on forty product cards and show the pass rate is more useful than one who knows a list of tricks. An analyst who turns a vague request into success criteria and a test set already does the core of the work. The evidence matters more than the title.

Starter skills and a portfolio of tested prompts

Clear writing comes first. Anthropic’s best practices suggest treating the model as a brilliant but new employee who lacks context, with a golden rule: show the prompt to a colleague who knows little about the task, and if they would be confused, the model will be too. Anthropic recommends three to five examples; Google’s guide warns that too many may make the model copy them. The other starter skills are plain: knowing the domain well enough to write the right answer yourself, keeping test cases in a spreadsheet and noting which model version produced each result.

A portfolio for this role is not a gallery of clever prompts. Each piece should show the task, the success criteria, the test set with its awkward cases, every prompt version with its score, and what still fails. Three such pieces from real work, such as a support answer template or a product description generator, say more than thirty screenshots of chat replies. Remove client names and personal data first.

One week of solo practice, then the AI course of VITON13 SCHOOL

This week, take one repeated task from your job and describe the good result as a rule a stranger could check. Collect twenty real inputs, add five awkward ones, and write down the answer you would accept for each. Draft a prompt, run all twenty-five, mark each result pass or fail, change one thing and run again. By Friday you will know whether the task suits a model and how much checking it needs, which is the core of what a prompt engineer does.

Alone, it is easy to stop at the prompt and skip the tests. A structured course adds an order of topics and feedback on your work. The AI course of VITON13 SCHOOL teaches writing prompts, automating routine tasks with AI and building simple AI agents; its page carries the current programme and terms, and the free school library is a place to start before deciding. For a website of your own rather than a new skill, see the guide comparing AI website builders with a web studio.