Assessment design
Is a Prompt Engineering Certificate Worth Anything, or Should You Give a Work Sample?
Give the work sample. A general prompt engineering certificate proves a course was finished on a date. It says nothing about whether the holder can frame a problem in your field, or knows what a claim there needs behind it. Two narrow uses survive: a tiebreaker between two juniors with no other evidence, and self-directed effort from a career-changer. A floor on a named tool belongs to a vendor certification. Never a filter. Forty-five minutes on your own material, graded against a key you wrote first, outranks any certificate.
The takeCertificates are what a hiring team reaches for when reading the actual work costs more attention than it wants to spend. That is the trade, and it is worse than it looks: the receipt is free to read and tells you almost nothing, while forty-five minutes of watching someone work costs forty-five minutes and tells you most of it. Most likely the credential also thins as it spreads, so any filter built on one ages faster than a task you could write this afternoon. What transfers is a field's own standard of proof, and no course sells that. It belongs to the occupation, not the tool.
Where Olive fits
Open a role and see what the work shows
The part a general certificate cannot carry is the field: what settles a claim in accounting is not what settles one in journalism. Olive authors its cases per occupation against that occupation's SOC code, and a human reviewer writes six separately-evidenced findings from the session, which is the same document the candidate receives.
Rank your shortlistIs a prompt engineering certificate evidence of anything?
It proves course completion on a date, about a tool. That is a real fact and a small one. It does not tell you whether the holder frames the problem before generating anything, whether they know what a claim in your field needs behind it, or whether they would catch the assistant's confident wrong answer. Those three decide whether AI help produces work you can ship.
The transfer problem is the whole story. When Sackett and colleagues revisited the meta-analytic validity estimates for selection procedures, the ones that came out on top were the job-specific ones (structured interviews, job knowledge tests, work sample tests, empirically keyed biodata), all ahead of general measures of psychological constructs 1. A general certificate is a general measure. It sits on the wrong side of that line by construction, and course rigor does not move it across.
What does transfer is smaller than the course pages suggest, and it depends on your field. A certificate holder has probably met structured output, iterative refinement, and the idea that a model will state something confidently with nothing behind it. That is useful. None of it tells them a comparable-company multiple has to come out of a filing rather than a summary, or that a denial appeal has to quote the payer's own policy language. That knowledge is occupational, and it is exactly where AI-assisted work goes wrong.
If the certificates in front of you are the big vendor ones, the specifics differ enough to matter: what each of those certificates actually requires separates a self-paced completion from a proctored exam. That changes how much weight the line deserves. It does not change what the line is evidence of.
Which credentials are worth anything, in order?
Ranked by what they let you conclude: a traceable artifact with history, then a reference who watched the person work, then a vendor certification for a tool your team actually runs, then a general prompt engineering certificate. Every one of them is weaker than forty-five minutes of that person doing your work while you keep the record.
| Credential | What it establishes | What it cannot establish |
|---|---|---|
| Traceable artifact with history | Work of this shape exists and this person's name is on it | Which parts they did, and which the assistant did |
| Reference who watched the work | How the person behaved over months, from someone with some incentive to be accurate | Anything comparable across candidates |
| Vendor certification for a tool you run | A floor on operating that specific product | Whether they know when not to use it |
| General prompt engineering certificate | A course was finished | Field framing, evidence standards, everything above |
| Forty-five minutes of your own work | What they framed, sourced, kept, refused and checked | How they behave over months |
The order is not an insult to certificates. It follows from how specific each item is to the job you are filling, which is the axis the validity evidence actually separates on 1. A traced artifact sits near your work, a general certificate sits far from it, and the work sample sits closest because it is the work.
The bottom row is also the only row you control. The other four arrive as claims about the past, written by people with an interest in them, on material you did not choose. A work sample is a record you commissioned, on a case you picked, graded against a key you wrote. That is the difference between reading evidence and producing it.
Two rows get conflated constantly. A shipped artifact and an exercise in working with an assistant answer different questions: the first asks whether the output exists, the second asks who produced the thinking behind it. The first stopped implying the second some time ago.
When does a certificate actually earn its place?
In three narrow cases. As a tiebreaker between two candidates whose evidence is otherwise identical. As proof of self-directed effort from a career-changer with no relevant artifacts yet. And as a floor on a named tool your team runs, where the certification tests operating that specific product and operation is genuinely what you need.
Not one of them is a reason to require one. They are reasons to stop discounting a candidate for having one, which is a different act.
The tiebreaker case is honest and genuinely small. Two juniors with the same absence of evidence (no artifacts, no relevant references, similar schools), and one of them spent their own evenings on a course. That is one bit of information about self-direction, and one bit beats zero. Weight it accordingly: it decides which of the two you invite, never which one you hire.
The career-changer case is stronger than it looks. Someone leaving healthcare administration for analytics has no analytics artifacts, by definition. The certificate is not proof they can do the work. It is proof they did unsupervised work nobody asked them for, which is a real signal about month two.
The tool-floor case is the narrowest and the most defensible. If your team runs one platform and the certification tests operating that platform, you are looking at a job knowledge test with a vendor's name on it, and job knowledge tests sit near the top of the validity ranking 1. Keep that apart from a general prompt engineering certificate in your own head, because the two are not the same instrument.
What none of the three survives is being used as a filter. A credential worth a small nudge inside a file becomes something else entirely when it decides who is in the file at all.
Design the 45-minute work sample
Take one real deliverable from the role, cut it down to what fits in forty-five minutes, and hand it over with an assistant that will do all of it if nobody stops it. Write the answer key first: which claims in the packet are wrong, which source settles each one, which shortcut the model will offer. Then grade the record, not the polish.
The material matters more than the format. Use a source packet from real work: three quotes written on different terms, a memo whose central claim only settles when someone opens the document under it, a dataset that will not correct a wrong reading. The assistant has to be able to produce a fluent, confident, wrong answer from it. If it cannot, the exercise measures typing.
The Uniform Guidelines are the right design constraint here even if nobody ever sues you. A content-valid procedure should closely approximate an observable work behavior, or its product should closely approximate an observable work product, and the manner, setting, level and complexity should closely approximate the work situation 2. Two clauses follow, and both cut real tasks: content validity is not available for knowledge an employee will be expected to learn on the job, and it is not available for a procedure that purports to measure a construct. The regulation's own list names judgment 2. So do not test your internal tooling, and do not write a rubric line reading "shows good judgment."
Score acts instead. Four of them, written down before the first candidate sees the task:
- What they asked for first. The deliverable, or the shape of the problem.
- Which claim they demanded a source for. Any claim, or the one the recommendation rests on.
- What they refused. A direction the assistant proposed and they turned down, with the reason stated.
- What they checked outside the conversation. A number recomputed, a document opened, a figure that moved as a result.
Then spend ten minutes on the submission with them. Follow-up questions about their own work are the cheapest instrument in the whole process, because a person who did the thinking can account for a choice and a person who accepted an output cannot. And grade the record rather than the artifact: if you grade the finished deliverable, everyone passes now, which is the polished take-home problem stated from the other end. See how Olive measures this
What happens if you screen on the certificate anyway?
You have made it a selection procedure. Under the Uniform Guidelines that is any measure used as a basis for an employment decision, and the definition names educational and training requirements and unscored application forms explicitly 3. So a certificate filter picks up the obligations of a test, on evidence a course completion cannot supply, and it cuts the people who learned the work without buying the course.
The legal exposure is ordinary and worth stating plainly. A neutral procedure that disproportionately excludes people on a protected basis has to be job-related and consistent with business necessity 4, and for a certificate filter that showing would have to be that the course content is a necessary prerequisite to important work behavior in the role 2. Most general course content is not, and the price of the course is part of who ends up holding one.
The practical cost lands first and hurts more. Certificates are cheapest to acquire for people with spare time and spare money, and least likely to be held by someone who learned the same things on a job that never asked them to certify it. A filter on the credential removes the second group and keeps the first, which is the reverse of what you wanted when you wrote the requirement.
There is a version of the requirement that survives. State the capability rather than the credential: can take a claim out of a source packet, find what settles it, and say what changed. Written that way it is checkable in the work sample and in ten minutes of questions on the resume claim, and it does not turn a course receipt into a gate.
Common questions
Should a job posting require a prompt engineering certificate?
No. Requiring one turns a course receipt into a selection procedure. The Uniform Guidelines define that as any measure used as a basis for an employment decision, and the definition names educational and training requirements outright. You would then owe evidence that the course content is a necessary prerequisite to important work behavior, which few general courses can support. Write the capability into the posting instead, and test it: framing the problem before generating, sourcing the claim the recommendation rests on, checking one thing outside the assistant.
Is it worth a candidate getting one?
For a career-changer with no relevant artifacts, yes: a general prompt engineering certificate is cheap evidence of self-direction and fills an otherwise empty line. For anyone who has work they can show, the same hours go further on one small artifact with a visible history, because a traceable piece of work answers a question the certificate does not. The exception is a certification that tests operating a specific platform a target employer runs. That behaves like a job knowledge test and carries more weight than a general course.
How long should the work sample be?
Forty-five minutes for the task, plus ten minutes on the submission. Long enough that the assistant can produce something wrong and the candidate has room to catch it, short enough that finishing it is not a favor. A longer task mostly buys a bigger deliverable, and the candidates who can afford the extra hours are not the ones you were trying to find. Cap it, say the cap in the invite, and grade the record of the work rather than the size of the output.
What does a vendor certification prove that a general certificate does not?
A floor on operating one specific product. If your team runs that product, the certification behaves like a job knowledge test, and job-specific procedures of that kind sit above general measures in the validity evidence. It still says nothing about whether the holder knows when not to reach for the tool, or what a claim in your field needs behind it. Treat it as one line in a file: a floor worth having, never a gate, and never a substitute for watching the person work.
Can you screen on the certificate and run the work sample later?
You can, and the screen is the part that costs you. Anything removing candidates before the work sample has to carry its own justification, and a course receipt cannot. Run it the other way: screen on something job-related and cheap, such as one traceable artifact or two questions about a claim on the resume, then spend the assessment on the shortlist. The certificate can sit in the file the whole time as context, informing your read and gating nobody.
References
- 1. Revisiting the design of selection systems in light of new findings regarding the validity of widely used predictors ✓ doi.org Job-specific predictors (structured interviews, job knowledge tests, work sample tests, empirically keyed biodata) top the revised validity ranking and beat general measures of psychological constructs; structured interviews lead at a mean validity of .42.
- 2. 29 CFR 1607.14 — Technical standards for validity studies (Uniform Guidelines on Employee Selection Procedures) ✓ ecfr.gov A content-valid procedure must closely approximate an observable work behavior or work product with manner, setting, level and complexity close to the work situation; content validity is unavailable for knowledge an employee will learn on the job and for constructs such as judgment.
- 3. 29 CFR 1607.16 — Definitions (Uniform Guidelines on Employee Selection Procedures) ✓ ecfr.gov A selection procedure is any measure used as a basis for any employment decision, explicitly including training programs and physical, educational and work experience requirements, and unscored application forms.
- 4. Employment Tests and Selection Procedures ✓ eeoc.gov A neutral selection procedure with a disproportionate exclusionary effect must be job-related and consistent with business necessity.
4 sources, numbered by first appearance. Every one was opened and checked against the claim it carries. 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.