Prompt Engineer Job Description Template 2026

Last updated: July 15, 2026

A strong prompt engineer job description makes clear the job is building and measuring LLM-powered systems, not writing clever prompts. It names the models and the evaluation tooling, and it posts a band, because the same title covers a $110K content-ops seat and a $400K research one. The template below comes from prompt and applied-AI searches that actually closed, not the wishlist reqs that pull 400 applicants and hire nobody.

Say “prompt engineer” out loud in a hiring meeting and half the room pictures someone typing questions into ChatGPT until the answer looks good. That picture is why so many of these searches go sideways. The typing is real. It is also maybe a fifth of the work.

The rest is measurement. A working prompt engineer spends most of their week deciding whether a change to a prompt, a retrieval step, or a model actually made the output better or just made it different, and proving it with numbers instead of a good feeling. That is a discipline. Most job descriptions for the role never mention it.

I’m Mike Carter. I lead partnership success at KORE1, which in plain terms means I sit between the hiring managers opening these reqs and the recruiters on our desk who fill them. We place prompt and applied-AI engineers through our prompt engineer staffing practice, part of our broader IT staffing services work, so yes, I have a stake here. We earn a fee when you hire through us. Bias noted. The template and the numbers below still work exactly the same whether you hand us the search or run it solo with a dozen tabs open at 11pm.

Scope the posting around the real work and the right people raise their hands. Scope it around “must be an expert prompter” and you get a slate of hobbyists who have never shipped a model into production. The posting is your first screen whether you designed it to be or not. So write it on purpose.

Prompt Engineer vs AI Engineer vs ML Engineer vs Data Scientist

A prompt engineer gets reliable behavior out of an existing model through prompting, retrieval, and evaluation; an AI engineer builds the application and infrastructure around that model; an ML engineer trains and deploys models; and a data scientist answers questions with statistics and experiments. Four adjacent seats, four different hires.

This is the boundary that sinks more prompt engineer searches than any skills gap. Draw it wrong on the posting and your shortlist fills up with people who are excellent at a job you did not mean to advertise. Happens a lot.

The prompt engineer works on top of models somebody else trained. Their raw material is the model’s behavior, and their job is to shape it: the instructions, the examples, the context you retrieve and feed in, the guardrails that keep it from going off a cliff, and the evals that tell you when it does. They rarely touch model weights. They touch everything about how the model gets used.

An AI engineer owns more of the stack. They stand up the API calls, the vector database, the orchestration, the caching, the fallbacks, the thing that pages someone at 3am when latency spikes. Plenty of people carry both titles at once, and at a small company one person genuinely does both jobs. Fine. Just fund it as two jobs when it is two jobs, and say so in the posting.

The ML engineer sits a layer deeper. They train, fine-tune, and serve models, and they live in PyTorch, GPUs, and the parts of the math a prompt engineer can mostly treat as a black box. If your bullets ask for training runs and model architecture, you are writing an ML engineering req and should pay for one. That is a scarcer, pricier hire.

Then the data scientist, who answers “what is happening and what should we do” with statistics, experiments, and models built to predict rather than to converse. Real overlap exists, because a good prompt engineer reasons about data all day. The line is that a data scientist’s product is an insight or a prediction, while a prompt engineer’s product is a feature that talks to a user and behaves.

Read the “core question” row before you write anything. If your honest answer is “train a model that fits our data,” that is an ML engineer. If it is “stand up the whole AI service and keep it running,” that is an AI engineer, or two of them. The prompt engineer is the person who makes the model do the right thing, most of the time, on inputs you have never seen, and can show you the score to prove it. Name that job. The right people answer.

What a Prompt Engineer Actually Owns

Under the noise, the work sorts into three layers. A posting that says which layer carries the weight pulls a sharper pool than one that lists every LLM buzzword in circulation and hopes the right person recognizes themselves in the pile. It rarely works.

The prompt and context layer. This is the part everyone pictures. Writing and versioning the instructions, choosing the few-shot examples, deciding what goes in the context window and in what order, and building the retrieval so the model sees the right documents at the right moment. It is real craft, and it rewards someone who is precise about language and stubborn about edge cases. Genuinely hard. It is also the layer people most overweight when they write the JD.

Because here is the layer that actually separates the strong hires.

The evaluation layer. The load-bearing one. A prompt engineer who cannot measure their own work is a prompt engineer who is guessing, and you will not find out until a customer does. Painful lesson. This is building the golden set of test cases, writing the graders, tracking hallucination and refusal rates, catching the regression when a model update quietly breaks a workflow that shipped fine last month. Tools like promptfoo, LangSmith, and Braintrust live here. When our recruiters screen for one skill that predicts whether someone will be good in the seat, it is this one. Just one. Can they prove a change helped, or do they just believe it did?

The integration and systems layer. Prompts do not run in a text box. They run inside a product, with a budget, a latency ceiling, and a pager. Real limits, every one. This layer is the API plumbing, the guardrails and content filters, the fallback when the primary model times out, the cost math when a clever prompt triples your token spend. A prompt engineer does not have to own all of it. They do have to think in it, because a prompt that is beautiful in a notebook and ruinous in production is not a good prompt.

One instinct runs under all three. Knowing what “good enough” means for your specific use case and being willing to say when the model is not there yet. A support bot that is wrong 4 percent of the time might be fine. A medical-summary tool that is wrong 4 percent of the time is a lawsuit. The engineer who can tell the difference, and who builds the evals that hold the line, is the one worth paying for. That judgment is the hardest thing to screen for on a resume, which is exactly why the interview has to grade the work. Not the trivia.

The Prompt Engineer Job Description Template

Here is the block. Copy it, swap the bracketed prompts for your real scope, and delete the italic notes before it goes live, because those are for whoever fills it in, not for the candidate reading it. It assumes a mid-level applied prompt engineer on an existing product team. Turn the ownership language up for a senior or lead, and down for a first AI hire who will be building the practice from a blank repo. Adjust from there.

Where Prompt Engineer JDs Go Wrong

I read a lot of these before a search opens. The same handful of mistakes shows up over and over, and each one quietly bleeds the exact candidates the hiring manager wanted. Here are the five I flag most.

The whole posting is about writing prompts. Not one word about measurement. So it attracts people who are clever with words and cannot prove a thing, and you do not find out until a model update breaks a feature nobody was testing. Fix it by naming the eval work loudly. Some of the best reqs I have seen put “evaluation” right in the title. Do that.

The opposite mistake costs just as much. A req asks someone to wire an LLM into a support flow, then demands a PhD and published research to do it. Now the applied builders who would own that work assume the bar is a doctorate they never earned, and they never apply. The researchers who do apply get bored of integration work by the second sprint and leave. You wanted a builder. You advertised for a scientist. Pick one.

Then there is the frontier-lab panic. Somebody reads that Anthropic and OpenAI clear $500,000, and either quotes a range they cannot honor or, worse, posts no range and hopes. Those are equity-heavy research seats at the companies training the models. They are not what it costs to wire an existing model into your product. Benchmark against applied roles at companies your size. That number is real, and you can actually pay it. Honest bands close.

Fourth, the tool salad. LangChain, LlamaIndex, promptfoo, LangSmith, Braintrust, Pinecone, Weaviate, pgvector, “experience with agentic frameworks,” all stacked under required. A strong engineer does not read that as a high bar. They read it as a team that has not decided what it is building. They bounce. Name the two or three tools the hire touches in week one. Drop the rest to preferred.

Last one is the simplest to fix and the most common. No band. In this market that reads as either disorganized or evasive, it burns the applicants who will not interview for a mystery number, and it wastes the screens you do land when the range finally surfaces and never worked in the first place. Four states make you post it now. List it.

Models, Tools, and Signals Worth Naming

Naming the real stack does two jobs. It filters for people who have actually used it, and it tells a strong engineer your AI work is a practice and not a slide in a board deck. Vague costs you both ways.

Start with the models. GPT from OpenAI, Claude from Anthropic, Gemini from Google, and open-weight options like Llama and Mistral all behave differently under the hood, and an engineer who has lived in one holds real opinions about the rest. Self-hosting an open model for cost or privacy is its own skill, separate from calling a hosted API, so if that is your setup, say it. The candidate who has done exactly that will recognize themselves in the posting and apply. Name your models.

The evaluation and orchestration tooling is where the posting gives itself away. Write “promptfoo, LangSmith, or Braintrust” and a serious candidate knows a serious team wrote it. For the plumbing, LangChain and LlamaIndex handle most orchestration, and Pinecone, Weaviate, or plain pgvector cover retrieval. None of those tools are sacred, though. Someone fluent in one eval framework learns the next in a week, so treat the specific product as transferable and the instinct to measure as the thing you cannot teach on the job.

The strongest signals are not tools at all. The person who talks about failure modes before features. The one who asks, unprompted, what error rate you can live with. The candidate with a real story about a prompt that dazzled in the demo and fell apart in production, and exactly what they changed. Certifications barely exist for this yet, and the few that do tell you almost nothing. A shipped feature with a number attached beats any credential on the page. Trust the receipts.

Prompt Engineer Salary Benchmarks for 2026

Most U.S. prompt engineers earn a base of $95,000 to $206,000 in 2026, with a national average near $129,500 across public aggregators and applied roles at normal companies clustering around $135,000 to $185,000. Frontier-lab packages at Anthropic and OpenAI clear $500,000 in total comp once equity lands, which is a different job than the one most teams are hiring for.

The sources agree on the middle and come apart at the edges, and the edges are where a hiring manager gets the band wrong. Right there.

Now look at the spread, because this is where the title lies to you. The government’s nearest proxy, Software Developers, sits at a $133,080 median with 15 percent projected growth through 2034 and roughly 129,200 openings a year. The aggregators cluster in the same neighborhood for a normal applied role. That is your tier. Then the frontier-lab numbers land two or three times higher, and they are real, but they are pricing a research seat at a company that trains the models, with a graduate degree usually expected and equity doing most of the work. That is not the job you are posting. When you set your band, benchmark against applied engineers at companies your size and treat the lab comp as a different market entirely.

Here is the part only a recruiter sees from the inside. Across the applied and prompt roles KORE1 closed over the last two quarters, the median base landed at $148,400, with the middle half between $118,000 and $192,000. We fill AI and IT roles in 17 days on average across more than 30 U.S. metros, and 92 percent of those placements are still in the seat a year later. Location still moves the number, though less than it used to now that the work is mostly remote. Still matters, though. A contract prompt engineer usually carries a 10 to 15 percent rate premium over the salaried equivalent, worth pricing in when you weigh contract or contract-to-hire against a direct hire. For the full breakdown by level and archetype, our prompt engineer salary guide goes deep.

Adapting the Template by Company Stage

The block above is a skeleton. Where your company sits decides which parts carry the weight.

Your first AI hire. If nobody has held this seat, the job is half building features and half building the practice. Say that plainly in the posting. This person sets up your first eval harness, decides which model you standardize on, writes the conventions everyone else will inherit, and has to earn trust with skeptical engineers who suspect the whole thing is hype. That takes a specific temperament. Comfortable with a blank page, patient with the doubters, allergic to shipping something they cannot measure. Rare mix. Lead with the mandate and name the executive backing it, because a first AI hire with no air cover stalls, and the good ones sense that setup in the interview.

An established AI or platform team. If there is already an AI group, a model of record, and an eval pipeline, the role narrows to a specific gap: a feature that keeps hallucinating, a migration to a new model, a RAG system that retrieves the wrong thing. Name the gap. Name the tools already in place and the domain the engineer steps into. A strong candidate reads that specificity as proof the practice is real. A generic AI post reads as proof it is not, and they keep scrolling. So get specific.

What Hiring Managers Ask Us Before Posting a Prompt Engineer Role

Is prompt engineering even a real job in 2026, or is it merging into AI engineering?

The title is consolidating; the work is not going anywhere. Roles that read “prompt engineer” two years ago now post as “applied AI engineer” or just “AI engineer,” and the day-to-day still lives in prompting, retrieval, and evaluation. Do not fight the label. Write the responsibilities you need, then choose whichever title pulls the right pool where you hire. Getting reliable behavior out of a model, and proving it with numbers, is a skill that keeps getting more valuable no matter what the recruiter calls it this quarter.

Do we need someone with a machine learning background?

For most applied roles, no, and requiring it is how these postings quietly strangle their own pipeline. Wiring a model into a product, shaping the context, building the evals. That is a software and judgment job, not a training job. A builder with solid Python and real shipping scars will out-deliver a researcher who has never owned a feature in production. The one exception is a role that actually fine-tunes or trains models, where you want an ML engineer instead, and that is a scarcer and pricier hire you should scope and pay for as such.

How do we actually test prompt engineering skill in an interview?

Hand them a broken prompt, real inputs, and forty-five minutes. Then watch the method, not the wording. Do they build test cases, define what “better” means, and measure the change? Or do they just nudge the phrasing until the one example on screen looks nicer? The first person is an engineer. The second is a hobbyist. Add a short design chat on how they would evaluate and guardrail the thing at scale, and skip the trivia round where you ask them to name frameworks. Frameworks are googleable. Judgment is not.

What separates a prompt engineer from an AI engineer, in one line?

A prompt engineer makes the model behave; an AI engineer builds and runs the system around it. At a small company that is one person. The distinction earns its keep when you decide what the hire prioritizes, so point the posting at whichever pain is actually bleeding, quality or uptime, and admit it if the honest answer is both.

Should we post the salary range?

Every time, no exceptions. Four states already force it, and everywhere else a missing band now reads as disorganized or evasive. This role earns an extra reason on top of the usual ones. The title stretches from a $110K content seat to a $400K lab seat, so the number itself tells applicants which job you actually mean before anyone burns a call finding out. Hide the band and the only thing you protect is the option to underpay. Strong engineers assume that is the reason and move on.

Contract or direct hire for a prompt engineer?

Follow how bounded the work is. A defined build, a proof of concept, or one feature with a real deadline points to contract staffing, which runs a 10 to 15 percent rate premium over salary and lets you start before the headcount fight is even settled. If the engineer will own a standing piece of how your product uses AI, hire direct. Not sure yet, and nobody on your team has managed this kind of work before? Contract-to-hire lets you watch the real output first. We place prompt and applied-AI engineers under all three, and the right one usually matches the shape of the work, not the box the req wants to check.

Next Steps

Take the template and make it yours. Pick the one shape you are hiring, name the models and the eval tooling, put measurement where nobody can miss it, split must-have from nice-to-have without flinching, and post the band. A short, specific prompt engineer JD beats a long generic one in every search we run. Every time.

Want a second read on a posting that keeps pulling the wrong slate, or help telling a prompt engineer role apart from the ML or AI-infrastructure job hiding inside it? That is our desk. Reach out to a recruiter on our team. KORE1 places prompt, applied-AI, and AI and machine learning talent across 30+ U.S. metros through direct hire, contract, and contract-to-hire, with an average time-to-fill of 17 days and 92 percent of those hires still in seat a year later. When you are ready to run the search, the full guide to hiring a prompt engineer walks it end to end, and the prompt engineer salary benchmarks break the comp down by level.