Roles we recruit
Seven practices, 27 functions. Start at the practice when you cannot yet name the title; go straight to the function when you can.
Browse all 229 roles Brief us on a seatFlat 15% of first-year base · shortlist in 10 working days
A prompt engineer designs, tests and maintains the instructions, examples and checks that make a language model’s output consistent enough to put in front of customers. Clever wording is easy to show, so we screen for evaluations they built, failures they caught before launch, and how they work with the people who own the product.
A new prompt engineer is judged on whether output quality becomes something the team measures instead of argues about. Each stage below has a signal that predicts it, and we test for that signal before you meet anyone.
The ramp, and the test behind each milestone
Weeks 1 and 2
A few hundred real outputs, customer messages or drafts, sorted into the ways they go wrong, with conversations with the people who use or review the results.
On track looks like a written list of failure types with examples and how often each occurs, agreed with the team that owns the output.
How we test for it before you meet them
A review exercise in our screen: twenty model outputs against a short brief. We score which failures they name first and whether they ask what a good output looks like.
Weeks 3 to 6
A fixed set of inputs, including hard and awkward ones, with a scoring method that is either a rubric a person applies or an automated check, and a baseline score for what runs today.
On track looks like a test set the team accepts as fair and a baseline score for the prompt in production.
How we test for it before you meet them
We ask for an evaluation they built: what it covered, how they checked the scoring and what it caught that reading samples missed. A former engineering or product lead confirms it.
Weeks 7 to 10
Instructions and examples revised one change at a time, each measured on the test set, with version history and a way back, so a regression shows up before release.
On track looks like a measured improvement on the baseline and a change log that explains it.
How we test for it before you meet them
We ask about a prompt change that made output worse somewhere unexpected, and how they found out. Strong answers name the test that caught it.
Weeks 11 to 13
Handling for bad or hostile inputs, checks before output reaches a customer, the cost and delay of each call, and a human review routine where a wrong answer is expensive.
On track looks like a documented review routine and a cost per task the team tracks.
How we test for it before you meet them
A reference call about a launch they supported: what they tested, what they blocked and what failed after release.
After day 90
Strong prompt engineers keep a library with owners and versions, rerun the tests when the model or the data changes, and teach colleagues to write prompts others can use.
On track looks like a library the team uses, with tests that run when the model version changes.
How we test for it before you meet them
We ask what they do when the model provider updates a model. A plan with tests and a way back predicts someone who maintains quality, not only builds it.
Five responsibilities. If the seat mainly writes code that calls models, it is an AI automation engineer; if it mainly decides which processes to automate, it is an AI workflow designer.
The instructions, examples and formats that tell a model what good output looks like for each task.
Fixed inputs and a scoring method, by person or by script, so quality is measured instead of judged by eye.
Reading real outputs, naming the failure types and deciding which a prompt can fix and which it cannot.
Controls on output that reaches customers, and a human review routine where a wrong answer costs money.
Versioned prompts with owners, a change history and tests that run when the model or the data changes.
Three neighboring seats absorb work people assume belongs here. Drawing the line wrong is how a prompt engineer ends up wiring integrations or choosing which processes to automate.
That seat ownsCustom automations in code: model and API calls, lead enrichment, research agents and data cleanup.
Hire it instead whenthe work is building and running the automation, with the prompts a small part of it.
That seat ownsMapping how work is done today and choosing which steps to automate, with the human checkpoints.
Hire it instead whennobody has decided which processes are worth automating first.
That seat ownsContent production with AI in the loop: workflows, brand and accuracy review, and output volume.
Hire it instead whenthe problem is a content workflow and its review, beyond the quality of one prompt.
The title stays the same; the work does not. Pick the one that matches what the model’s output feeds.
Owns the prompts behind a feature customers use, such as a support assistant, a summarizer or a copilot.
Builds prompt systems for content, campaigns and outreach, with brand voice, accuracy and review built in.
Specializes in test sets, scoring and regression checks for model output, often across several teams.
The questions we ask in our own screen. Use them in yours; the warning signs are as useful as the answers.
Every candidate we put forward is scored on these five, and the scores travel with the shortlist, gaps included. Four out of five with a named weakness beats a clean sweep you cannot verify.
| What we score | Weak evidence | Strong evidence |
|---|---|---|
| Evaluations they built | Tries examples by hand. | A test set, a scoring method and a failure it caught, shown. |
| Failure analysis | Adds instructions until it passes. | Sorts real failures into types and fixes the cause of each. |
| Change discipline | Edits live prompts with no history. | One change at a time, versioned, measured and reversible. |
| Judgment about limits | Treats every problem as a wording problem. | Knows when retrieval, data or a human step is the fix. |
| Working with the owners | Delivers prompts only they can maintain. | Prompts and tests the product team can read and change, confirmed by that team. |
Public US benchmarks, read October 2026. Neither Salary.com nor Built In has a benchmark page for the title, so the chart uses the two sources that publish it, which measure different things, and sets the nearest AI seat and the BLS catch-all for computer jobs beside them. Price on scope: a seat that owns evaluation for a customer-facing product sits well above one that drafts prompts for a team.
Each bar runs from the 25th to the 75th percentile, the middle half of reported pay.
Pay range, 25th to 75th percentileWages including commission (BLS)Average base salary from job postings (Indeed)Average total pay, base plus variable
Prompt Engineer
Indeed, average base from job postings
$112,522 average base salary47 job postings
Prompt Engineer, total compensation
Levels.fyi, self-reported
$80k to $268k total compensation$170,000 median, 19 reports
AI Engineer
Salary.com base, Built In total, neighboring seat
$107,833 to $125,338 base$211,243 average total
Computer occupations, all other
BLS, all industries, jobs not listed separately
$79,370 to $157,500
Prompt Engineer dot: Indeed average base salary from 47 job postings, United States, updated October 11, 2026. Prompt Engineer bar: Levels.fyi total compensation, 25th to 75th percentile, self-reported by 19 people, all levels, last updated October 19, 2026. AI Engineer bar: Salary.com base salary, 25th to 75th percentile, United States, as of October 1, 2026; ring: Built In average total compensation (base plus additional cash, self-reported), reviewed October 18, 2026. Computer occupations, all other: U.S. Bureau of Labor Statistics, Occupational Employment and Wage Statistics, May 2025, annual wage at the 25th and 75th percentiles (SOC 15-1299), all industries. Public aggregates, not ZaphyrRecruit placement data.
A prompt engineer designs, tests and maintains the instructions, examples and checks that make a language model’s output consistent. The daily work is reading real outputs, naming how they fail, building a test set and scoring method, changing prompts one step at a time and watching for regressions when the model or the data changes. In a go-to-market company the output is often content, outreach, support answers or research.
Indeed’s average base salary is $112,522 from 47 job postings (October 2026). Levels.fyi, which reports total compensation, shows a $170,000 median from 19 reports, with the middle half between $80k and $268k. Salary.com puts AI engineers at $107,833 to $125,338 base. The spread is wide because the title covers everything from drafting prompts to owning evaluation, so price the seat on scope.
If your problem is building and running the system, an AI engineer or automation engineer is the hire. If the system works but output varies from excellent to unusable and nobody measures it, you need someone whose job is evaluation. A prompt engineer for hire on contract suits a first test set; in-house pays off when output reaches customers every day.
Models keep changing, and wording tricks that work on one version often fail on the next. The lasting work is deciding what good output is, testing it and catching regressions when a model or the data changes, and that needs an owner. We screen for evaluation judgment rather than prompt tricks, because it still applies after the next model release.
Clear writing, enough data skill to build and score a test set, and enough technical knowledge to read model settings, logs and API behavior. The skill that separates candidates is evaluation: sorting failures into types, changing one thing at a time and knowing when a prompt is the wrong fix. Domain knowledge of your content or customers helps more than a certificate.
Give them twenty real outputs and a one-paragraph brief, and ask what is wrong, what they would test and what they would change first. An hour is enough. It shows how they sort failures and whether they ask what good means, which a portfolio of clever prompts hides. Ask for an evaluation they built and call someone who used it.
Ten working days from a signed brief to a shortlist of three to five calibrated candidates, each with our output review exercise scored and an evaluation they built checked with a former colleague. Interview speed decides the rest: strong candidates rarely run one process at a time, so a decision within two weeks of the first interview keeps them.
Tell us which product feature or workflow the output feeds, what goes wrong today and who reviews it. We come back within one business day with how we would run the search and what the seat pays.
We build shortlists from people doing this work now, often before a seat is advertised. Set up a job alert for AI roles, or send your CV and we will come to you when a seat opens.
The prompt engineer is one of twelve seats in this function. If the gap sits elsewhere, start there.