Design take-home assignments

The failure nobody measures

The candidate who spends fifteen hours on your assignment is usually the one with the fewest competing offers. The candidate with three live processes declines politely, and you never learn how good they were.

This filter leaves no trace in your funnel. You see a shortlist and assume it represents the market, when it partly represents who had a free weekend.

Uncomparable results push you towards effort

Two candidates spend different amounts of time, choose different scopes and present in different formats. When you sit down to decide, the only genuinely comparable dimension left is how much work each one produced.

That is how teams end up rewarding polish — not because they value it, but because it is the one thing the format made comparable.

A finished deliverable cannot show adaptability

The brief a candidate receives is the brief they keep. Nothing is ruled out halfway, no deadline moves, no stakeholder pushes back.

So the single most predictive behaviour for senior work — what someone does when the plan breaks — is structurally invisible in this format, no matter how good the prompt is.

If you are keeping your take-home, fix these four things

Most of the damage comes from format, not from the questions. These are cheap to change.

• Cap it at two hours and say you will only read two hours of work. Then actually stop reading.
• Give every candidate the identical scope, with identical data. Comparability is the whole point.
• Write the rubric before you send it, and share it with the candidate. If you cannot write the rubric, the exercise is not ready.
• Pay for it if it takes more than two hours, or if it touches anything resembling your real roadmap.

Should take-homes be paid?

If the exercise is genuinely fictional, short, and produces nothing usable, payment is not the issue — time is. Cap it instead.

If it takes a full day, or if it is recognisably your product, then yes, pay for it. And be aware that a paid full-day exercise still filters out the candidates with options; payment fixes the ethics, not the sampling problem.

What a take-home is genuinely good at

Everything above is a criticism, and a page that only criticises is not useful to someone who has to run a process next week. There are three things this format does better than anything else, and they are worth protecting.

It shows craft at full resolution. An interview shows you how someone talks about spacing and hierarchy; a take-home shows you whether they can actually do it. For a role where the output quality is the job, nothing substitutes for seeing the artefact.

It shows how someone communicates work to people who were not there. Most design work fails at that step rather than at the design step, and a take-home is a direct sample of it.

And it is the only format that shows what someone does with unbounded time. That is a real signal — some people expand to fill it and produce noise, others use it to go deeper on one thing. Just be aware you are measuring that, and not confuse it with ability.

The prompt matters more than the format

Most take-homes fail before the candidate opens the document, because the prompt asks for a deliverable instead of a decision.

«Design a booking flow for a fictional gym» produces six booking flows that all look competent and tell you nothing, because there was nothing to get wrong. Every candidate makes the same choices, and you end up comparing typography.

«Here is a booking flow we built. Here is the data showing where people drop out. You have two hours — tell us what you would change and what you would not touch» produces six different answers, and the differences are the information. One candidate rebuilds everything. One finds the actual drop-off. One says the data is not sufficient to conclude anything and explains what they would measure first.

The rule: if every good candidate would produce roughly the same thing, the prompt is not asking anything. Put a genuine trade-off in it — something has to be given up — and the answers separate immediately.

How to score it so two people agree

The rubric problem is not that teams do not have one. It is that the rubric has categories like «craft» and «thinking», which two reviewers will score three points apart and both be right.

Score against observable events instead. Not «did they show good judgment» but «did they state at least one thing they chose not to do, and why». Not «good communication» but «could a stranger understand the recommendation from the first paragraph alone». Those have answers.

Then check your agreement before you trust the scores. Have two people score the same three submissions independently. If the totals are more than a point apart on a five-point scale, the rubric is measuring the reviewer rather than the candidate, and adding more candidates will not fix it.

When to skip the take-home entirely

Three cases where the format is the wrong tool and something else is strictly better.

• You are choosing between six people. A take-home at that stage costs the candidates twelve hours between them and costs you six reviews, and most of them were never going to make it. Narrow first with something cheap, then use the take-home on two.
• The role is senior. Senior candidates decline unpaid exercises at a much higher rate — not out of arrogance, but because they have options and a job. You will get a shortlist made of the people with the least leverage.
• You cannot write the rubric. If you sit down to define what a good answer looks like and cannot, the exercise will be scored on impression, which is what you were trying to avoid by running it.

What to use instead, and what it costs you

The honest alternatives, with what each one gives up.

A live exercise — an hour, together, with the constraint changing halfway. Removes the time-availability problem entirely and shows you how someone handles a plan breaking, which a take-home structurally cannot. Costs you an hour of your own time per candidate, and it disadvantages people who think slowly and well.

A structured conversation about work they already did, with the same questions for everyone. Cheapest thing on this list, and much better than an unstructured chat. Gives up the ability to see anything they did not choose to show you.

A situational assessment — written scenarios, no correct option, scored the same way for everyone. Fifteen minutes of candidate time and fully comparable, which is the property take-homes lack. Gives up seeing the craft, so for a role where output quality is the job you still need eyes on real work at some point.

Frequently asked

How long should a design exercise be?

Under two hours if unpaid, and enforced rather than suggested. Beyond that you are sampling availability rather than ability, and your response rate from candidates with competing offers drops sharply.

Is it acceptable to use a candidate's exercise internally?

No. If the exercise touches your real product it stops being an assessment and becomes unpaid work, and experienced candidates will recognise it immediately and decline. Use a fictional scenario.

What is a reasonable response rate for a design take-home?

If more than a third of the people you invite decline or go silent, the exercise is too long or the process is too far along without a conversation. Track it — it is the cheapest diagnostic you have, and most teams never look at it because declines do not show up in a hiring funnel as a problem.

Should we tell candidates what the rubric is?

Yes. The fear is that they will optimise for it, and they will — which is fine, because the rubric describes the things you actually care about. What sharing it removes is the advantage held by candidates who happen to guess right about what you want, which is not a quality you are hiring for.

Can we reuse the same take-home for every candidate?

You should. Comparability is the point, and a rotating set of prompts destroys it. The risk of the prompt circulating is real but small, and it is a smaller problem than not being able to compare two candidates — if a prompt leaks, the answers still differ, because there was never a correct one.