If you're searching for a take-home interview alternative, you already suspect what I'm about to argue: the format is done. Not wounded, done. And the interesting part is that AI isn't really what killed it.
I run an assessment company, so I have an obvious bias here, disclosed up front. But I talk to engineering leaders about their screens every week, and the take-home conversation has the same shape every time. They stopped trusting the results a while ago. They kept the format because the alternative seemed to be going back to whiteboard trivia. This piece is about the third option.
What take-homes got right
Let me steelman the thing before burying it, because the take-home became popular for real reasons.
It was async. No scheduling tetris across time zones, no burning a senior engineer's Tuesday. The candidate worked when they wanted, in their own editor, with their own tools. That's closer to real work than any whiteboard session ever was.
It also respected a certain kind of fairness. People who freeze in live interviews got to show what they could actually build. The hiring-without-whiteboards list exists because thousands of engineers hated performative live coding, and take-homes were the popular refuge.
Whatever replaces the take-home has to keep those properties. Async, realistic, low theater. Hold that thought.
Then AI started doing the homework
Here's the mechanism people miss. A take-home is an unobserved artifact. You hand out a prompt, days pass, a repo comes back. You were never grading a person. You were grading a deliverable and assuming the person produced it in some particular way.
That assumption is now worthless. Ask engineering leaders what share of their take-home submissions are AI-assisted and you'll hear anywhere from a third to half, and the honest ones add that they can't tell which half. I can't verify those numbers and neither can they. That's the point. Nobody can, and a screen nobody can verify isn't a screen.
The usual response is to fight for the old assumption. Add an AI-use policy. Run detection on the submission. Interrogate suspicious commits. I wrote a whole piece on why detection is the wrong question, but the short version: you can't police what happens on someone else's machine over three days, and every detection method false-flags strong candidates while missing anyone who retypes output. The arms race is unwinnable on the defender's side.
So the take-home didn't die of AI, exactly. It died of unobservability. AI just raised the price of not watching.
The costs you were eating before AI showed up
Even when take-homes "worked," the bill was bigger than most teams admitted.
Drop-off, first. A four-hour unpaid assignment reads very differently to a candidate with three offers in flight than to one with none. The people who refuse are disproportionately the people you wanted. There's a long-running HN thread ranking real take-home tests where the recurring complaint isn't difficulty, it's disrespect for candidate time. Your funnel felt fine because the drop-outs were invisible.
Second, senior time on the grading side. Somebody has to review every submission properly, and "properly" means reading code from a stranger with no context. Teams either did it well and paid hours for it, or skimmed and pretended.
Third, and this is the one that stings: false confidence. A polished repo feels like strong signal, and confident wrong beliefs about candidates are more expensive than admitted uncertainty. But your bad hires were rarely people who couldn't produce a clean artifact. They were people who went quiet when stuck, never asked the obvious question, or couldn't take a code review without bristling. The artifact never carried any of that. The take-home was quietly measuring the narrow thing while everyone read it as measuring the broad thing.
If you must keep take-homes: the minimum-harm version
Some teams will keep them anyway this quarter, for inertia or tooling reasons. Fine. Then at least run the version that does the least damage:
- Two hours, hard cap, stated in the brief. Longer selects for free evenings, not talent.
- Pay for the time. It changes who's willing to do it and it changes how your company reads to seniors.
- Allow AI explicitly. It's happening anyway. Making it legal removes the charade and lets you ask about it honestly later.
- Put the real evaluation in a 30-minute debrief. Have them walk you through what they built, why, and what they'd change. Ask what the AI got wrong. Someone who did the work answers instantly. Someone who transported it stalls.
- Grade the debrief, not the repo. The artifact is now just the conversation piece.
Notice what happened: the fixed version moved all the signal into the part where you watch a human think. That's not a coincidence. That's the replacement announcing itself.
What actually replaces them: watching the work
The take-home's async, realistic core is worth keeping. The blindness isn't. Put those requirements together and you get an observed work simulation: a scoped slice of real work, done in about an hour, self-scheduled, AI tools allowed and visible, the whole session recorded, with a short defense of decisions at the end.
Here's where I tell you that's what I build. Skillvee is a 60-minute "day at work" simulation that replaces the recruiter phone screen and technical first round. Candidates solve a real challenge, talk to AI peers, and defend their decisions to an AI manager while the screen records. In one assessment you see how a candidate codes, communicates, collaborates, exercises agency, and leverages AI, before any senior engineer spends an interview hour. Bias fully disclosed. The comparison below still holds if you build something like this in-house.
How the three formats actually stack up:
| Take-home | Live coding round | Observed simulation | |
|---|---|---|---|
| What you grade | The artifact | Performance under watching | The working process |
| Candidate time | 3 to 8 hours, usually unpaid | About an hour, scheduled | About an hour, self-scheduled |
| Your team's time per candidate | 30 to 60 min grading | An engineer-hour plus prep | Minutes reviewing a scored report |
| AI reality | Invisible, unpoliceable | Usually banned, so unrealistic | Allowed and observable |
| What it misses | Everything about the person | How they work without an audience watching live | Long-horizon collaboration |
| Drop-off pressure | High, worst among strong candidates | Medium | Low |
| Scheduling | Async | Painful | Async |
The right column keeps the async and realism that made take-homes popular, and adds the one thing they never had: you see the person work. Communication, agency, judgment, and AI use stop being inferences from a repo and become things you watched happen. Those are the dimensions that predict whether the hire works out, and there's a longer breakdown of all six in the AI-cheating piece.
The debrief is the tell
One pattern worth stealing regardless of format. Whatever a candidate hands you, the highest-signal ten minutes you can buy is making them defend one decision under mild pushback.
Not a quiz. One question, sincerely asked: why this approach and not the obvious alternative? People who did the thinking enjoy that question. You can hear it, they lean in, because someone finally asked about the interesting part. People who outsourced the thinking treat the same question like a trap, and the stall is audible too.
In our simulations that moment is built in, an AI manager pushes back on one choice near the end, but you can replicate it in any process for free, today, with a calendar invite. If your current take-home has no defense step, you're grading fan fiction.
Where this gets hard
Honesty section, because switching isn't free.
A good simulation task is real design work. Too scripted and it's a fancy quiz. Too open and you can't compare candidates. The task needs deliberate ambiguity, a reason to ask questions, and a decision worth defending. Ours took many iterations to get right, and an in-house version will too.
Consistent scoring is the other cost. One person watching recordings on vibes drifts within a week. You need a written rubric your team actually follows, or tooling that applies one for you. And if your hiring managers won't engage with the output, whether that's recordings or scored reports, you've built signal nobody reads, which is worse than no signal because it feels rigorous.
If your volume is two hires a year, honestly, a paid two-hour take-home plus a serious debrief gets you most of the way. The simulation math starts winning when screening time compounds, which is roughly when you're hiring more than a couple of engineers a quarter.
The take-home had a good decade. It earned its spot by being more humane than whiteboards, and it lost that spot the day the artifact stopped proving anything. Grade the process now. It's the only thing left that can't be faked.