Skills testing means asking a candidate to do a piece of the actual work and assessing the result. It is consistently among the strongest predictors of job performance, for the unsurprising reason that it is the job rather than a proxy for it.
It is also the assessment most often designed badly: too long, too artificial, unpaid when it should not be, and scored on impressions rather than criteria. A good test takes an hour to design and pays for itself immediately. A bad one loses you your best candidates and tells you very little about the ones who stay.
Test the work, not a puzzle
The test should be a small, real version of what the person will do.
For a support role, a genuinely awkward customer email to respond to. For an analyst, a messy dataset with a specific question to answer. For a marketer, a brief and a draft. For a developer, a small change to a realistic codebase.
Abstract puzzles and brainteasers predict very little about job performance and irritate exactly the candidates who have options. If the work does not involve inverting a binary tree under time pressure, do not test for it.
The test should be answerable from the job description. If a candidate could not tell from the advert that this capability mattered, the test is measuring whether they guessed correctly.
Keep it short, and respect that their time is not free
The single most common design failure is length. A four-hour take-home task filters for people with free evenings, which correlates with having no caring responsibilities, no second job and no current employment. That is not the filter you intended.
Sixty to ninety minutes is the right ceiling for most roles. Below thirty minutes you usually cannot see much.
State the expected time honestly, and mean it. If you say ninety minutes and it takes four hours, you learn how long a candidate will spend rather than how good they are, and you have damaged your reputation with everyone who did it properly.
Pay for anything substantial. Any task beyond about ninety minutes, and certainly anything you could conceivably use, should be paid at a fair rate. This is increasingly expected, it is straightforwardly the right thing to do, and it makes candidates take the task seriously.
Never use the output. Do not set a task that solves a live problem. Candidates notice, it is exploitative, and it will eventually be posted publicly.
Live or take-home
Take-home suits work that needs thinking time and produces an artefact. It is fairer to people who interview badly but work well. It is harder to attribute, since you cannot be certain who did it or with what help.
Live suits work involving reasoning, conversation or reacting to new information. You see how they think and can ask why. It disadvantages people who freeze when watched, which is often unrelated to the job.
A pairing that works well for many roles: a short take-home, followed by a live conversation about it. "Walk me through why you approached it that way. What would you do differently with another day." Attribution ceases to be a concern, because someone who did not do the work cannot discuss it, and you get reasoning as well as output.
This also resolves most of the anxiety about AI assistance. If a candidate used a model to help, that is increasingly just how the work gets done. What matters is whether they can explain and defend the result, which is exactly what the conversation tests.
Score against criteria written in advance
Write the rubric before anyone takes the test. Three to five criteria, each with a description of weak, adequate and strong.
For the support email: did they resolve the actual problem, did they handle the emotion, is the tone right for our brand, is it clear and correct, did they escalate appropriately.
Score each criterion separately. A single overall impression is where bias re-enters through the side door, usually attaching to writing polish rather than substance.
Have two people score independently for important roles and compare. Disagreement means your rubric is ambiguous, which is worth knowing before you use it on twenty more candidates.
Score blind where you can. Strip the name from the submission. Work samples are one of the few stages where genuine blind assessment is easy, and it is worth doing.
Where to put it in the process
After an initial screen, before the main interview, for most roles.
Putting it first, before any human contact, produces heavy drop-off, because you are asking for an hour of unpaid work from someone who has had no signal that you are serious. Putting it last, after several interviews, means you have already spent the expensive time.
Tell candidates it is coming when the process starts, so nobody is ambushed by a task with a two-day deadline while managing a current job.
Off-the-shelf or build your own
Build your own where the work is specific to you and you can write a realistic task. Highest validity, costs you design and marking time.
Off-the-shelf where the skill is standard and testable: a language, a software package, a professional competency. Consistent, normed, no design or marking time.
HRM8's Assessment Hub provides library skill tests from US$25 per candidate, at three levels: foundational at US$25, specialist at US$35 and advanced at US$45. Curated role assessments bundle the relevant tests by seniority, from US$35 for entry level up to US$130 for executive roles. Everything is per candidate, credits do not expire, and results are AI-scored with no automatic rejection: the employer owns the decision.
The practical split most employers land on is an off-the-shelf test for the standard technical component and a short bespoke task for the part that is specific to their context.
Common questions
How long should a skills test take?
Sixty to ninety minutes at most for an unpaid task. Anything longer should be paid, and anything much shorter rarely shows you enough.
Should you pay candidates for a work sample?
Yes for anything substantial, certainly beyond ninety minutes. It is increasingly expected and it improves the quality of what you get back.
What if a candidate uses AI to complete the task?
Assume some will, since that is how the work is now done. Follow the task with a conversation about it. Someone who cannot explain and defend their approach has told you what you needed to know.
Where should a skills test sit in the process?
After an initial screen and before the main interview. First produces drop-off, last wastes your interview time.
Are work samples better than interviews?
As predictors of whether someone can do the work, generally yes. They are complementary rather than competing: the work sample shows capability, the structured interview shows reasoning and how someone operates.
Test a real slice of the job rather than a puzzle, cap it at ninety minutes unpaid, pay for anything longer, write the rubric before anyone takes it, score blind and independently, and follow a take-home with a conversation about the reasoning.