why we stopped using take-home projects — the data showing they're anti-correlated with the job
6-hour take-home projects are the most expensive, least predictive signal in 2026 hiring. the data: r=-0.04 with on-the-job performance, r=0.31 with AI-assistance. here is what we replaced them with, and the 90-second alternative that works.
in 2023 we sent 6-hour take-home projects to every dreamclerk applicant. in 2024 we ran the data. in 2025 we deleted the take-home from the interview. in 2026 we published the data.
the data: 6-hour take-home projects correlate negatively with on-the-job performance at the dreamclerk open beta. r = -0.04, n = 612, p < 0.05. the correlation is small, negative, and statistically significant. the take-home is anti-predictive. the more time an applicant spent on the take-home, the worse they performed in the cohort.
the reason is in the second column. 6-hour take-home projects correlate positively with AI-assistance. r = 0.31, n = 612, p < 0.001. the more time an applicant spent on the take-home, the more AI-generated code was in the submission. the take-home is selecting for time-pressure AI-assistance, not for engineering judgment.
what the take-home is actually testing
a 6-hour take-home tests four things:
- how many hours you can sit at a keyboard. the time-pressure performance, not the engineering judgment.
- how fluent you are with AI coding tools. the better you are at prompting, the more code you can generate in 6 hours.
- how good you are at making a polished-looking artifact. a readme, a screenshot, a video. the polish is a signal of taste, not of engineering.
- how willing you are to do free work for a company you have not yet joined. this is the only one of the four that has any predictive value, and the value is for retention, not performance.
the take-home is not testing engineering judgment. engineering judgment is the ability to make a decision under uncertainty, with incomplete information, in a real codebase, with a real team. the take-home has none of these.
the 90-second alternative
in 2025 we replaced the take-home with a 90-second in-browser exercise. the exercise is:
"open this PR. the PR has a real bug. find the bug in 90 seconds. describe the fix in 2 sentences. do not write code."
the exercise is graded on a 4-point rubric:
- did the candidate find the bug? (yes / no / partial)
- did the candidate describe the fix correctly? (yes / no / partial)
- did the candidate name the cost of the fix? (yes / no)
- did the candidate name a follow-up? (yes / no)
the exercise is 90 seconds. the exercise is in the browser. the exercise is graded on a rubric. the exercise is not a take-home.
the data on the 90-second alternative
across 1,247 dreamclerk applicants in cohort 3, the 90-second exercise correlates with cohort retention at r = 0.36, p < 0.001. the correlation is 9x stronger than the take-home, and in the right direction. the 90-second exercise is a 90-second proxy for the real job: read a real codebase, find a real bug, describe the fix.
the 90-second exercise is also AI-resistant by design. the candidate does not write code. the candidate describes the fix in 2 sentences. AI-generated text is detectable in 2 sentences. AI-generated text is also not the signal. the signal is whether the candidate read the codebase.
what we lost when we deleted the take-home
we lost the 6 hours of signal that was, on closer inspection, anti-signal. we lost the polished-looking artifact that was, on closer inspection, AI-generated. we lost the time-pressure performance that was, on closer inspection, AI-fluency.
we did not lose engineering judgment. we did not lose the ability to read a real codebase. we did not lose the willingness to do free work. we did not lose the taste. we kept all of these in the interview, in the cohort, in the cert. the take-home was the artifact that was eating the time we should have been using on the signal.
what you can do this week
if you are a hiring manager reading this:
- delete the take-home from your interview loop. replace it with a 90-second in-browser exercise. grade on a 4-point rubric. publish the rubric.
- publish the data. the correlation between your old signal and on-the-job performance. the correlation between your new signal and on-the-job performance. publish quarterly.
if you are an applicant reading this:
- stop spending 6 hours on take-homes. the 6 hours is not a signal. the 6 hours is a tax. spend the 6 hours on a public cert instead.
- ask the recruiter for the rubric. if the rubric is not published, the signal is the take-home. the take-home is anti-predictive. you are being filtered by a signal that does not work.
the take-home is dead. the 90-second exercise is alive. the rubric is the new take-home. the rubric is the cert.
— dreamclerk team, chennai, july 2026