How to Design a Take-Home Test That Actually Predicts

You've just read a CV that says "advanced Excel" and sat through an interview where the candidate talked fluently about pivot tables. You still don't know if they can build one. That's the gap a take-home test is supposed to close — and the gap most take-home tests fail to close, because they're either too easy to fake or too long for anyone reasonable to finish.

Here's how to build one that actually tells you something.

Start from the task, not the skill

Most take-home tests start with a skill ("test their SQL") and end up as a trivia quiz. Start with a task instead: the actual thing this person will do in month two on the job. If the role is a bookkeeper, the task is reconciling a messy set of transactions, not answering multiple-choice questions about debits and credits. If the role is a customer support hire, the task is drafting three replies to real (anonymised) angry-customer emails, not defining "empathy."

A good test question to ask yourself: could someone pass this by memorising an answer, or do they have to actually produce something? If it's the former, redesign it.

Keep it to the time you'd actually give them on the job

This is where most companies lose good candidates. A test that takes six hours signals that you don't respect people's evenings, and it filters for candidates with no other options rather than candidates with the best skills. A reasonable cap:

• Junior or entry-level roles: 45–60 minutes
• Mid-level individual contributor: 60–90 minutes
• Senior or specialist roles: up to 2 hours, split across two smaller tasks if you can

If your test genuinely needs more time than that, it's not a screening tool anymore — it's unpaid work, and candidates increasingly recognise it as such and walk away. State the time limit up front in the email, and mean it: "This should take about an hour. If it's taking you three, stop and send what you have — we'd rather see your instinct at the one-hour mark than a polished version at hour four."

Give them the same messy inputs the job gives you

Real work is rarely clean. If you're testing a data-entry or analyst role, don't hand over a spreadsheet with no errors in it — that's not what they'll get on day one. Include a duplicate row, a mislabeled column, an ambiguous instruction. How someone handles the mess tells you more than how they handle the parts you already tidied up for them.

For a writing or marketing role, give a real (or realistic) brief with a contradiction buried in it — a client who says "keep it short" and "cover everything" in the same paragraph. Watching whether they ask a clarifying question or just guess is often the most useful five seconds of the whole exercise.

Score it before you read a single submission

Write the rubric first, blind. Decide what a 1, 3, and 5 look like on each dimension before you've seen anyone's work — otherwise the first strong submission you read quietly becomes your definition of "good," and everyone after it gets judged against a moving target.

A workable rubric for most tasks has three to five dimensions, each scored 1–5:

1. Did they solve the actual problem (not a nearby, easier one)
2. Would this be usable with zero further editing from you
3. Did they show their reasoning, not just an answer — a comment, a note, a "here's why I chose this"
4. Did they handle the messy or ambiguous part well
5. (Optional) Did they respect the brief — length, format, tone

Have two people score independently if you can, even just you and one other person on your team, and compare notes before you talk about the candidate. You'll be surprised how often you disagree, and that disagreement is useful information about the rubric, not just the candidate.

Don't ask for free work you'll actually use

If the task is "redesign our onboarding email" and you then use the redesign, you've extracted free labour under the cover of a hiring process — and candidates talk about this in ways that follow a company for years. Two honest ways around it:

• Use a fictional or past scenario that has no live business value (an old campaign you're not running anymore, a made-up client).
• If you genuinely want to test on live work, pay for the assignment, even a token amount, and say so upfront.

Where a written test still can't tell you enough

A take-home tells you what someone can produce alone, at their own pace, possibly with help from a friend or a search engine. It won't tell you how they perform under a bit of pressure, how they take feedback in real time, or whether the polished submission was really theirs. That's what the follow-up conversation is for — ask them to walk you through one decision they made and why, and watch how fluently they can defend it. Someone who built it will do this easily. Someone who outsourced it usually can't.

If you're building one good take-home test for one role, writing it yourself the way above is the right amount of effort. If you're hiring across five or six different roles a month and rewriting a rubric every time is eating your week, that's the point where a calibrated testing platform earns its keep — AssessFit runs proctored, adaptive tests with a built-in integrity score so you're not starting from a blank rubric each time, and it's free to try up to five candidates a month.

Either way, the test is only half the job. The other half is being honest with yourself about what it actually measures — and not pretending it measures things it doesn't.