We moved from a take-home assignment to a 90-minute code comprehension and live refactor round, allowing AI tools. We want to test how well a candidate reads and understands code, and which principles and best practices they apply when structuring a small project.
Surprisingly, candidates seem to struggle more with this new interview.
LeetCode isn't going away because most interviews are bad at measuring everything else and comfortable measuring code recall. The real question is what replaces the pattern matching part when AI handles it, I think it is explaining your tradeoffs on a messy real system. Are you seeing companies shift to live design or incident-style questions yet?
System design interview is more important than ever, that's my feeling talking with different folks. Additionally, I do see a "Debugging an existing application" type of exercise. Where the candidate is given a working application with a hidden bug, and is asked to investigate it on the interview.
agree, I think the trend will be less leetcode style, more AI assisted. At least ten companies already run a coding round with an AI assistant in the room, up from one (Canva) in June 2025.
the debugging exercise you mention fits that well. In practice data from Nyrion (I run it, a practice platform for these rounds) people accepted AI’s code 93% of the time, and fewer than 1 in 10 answers explaining their own passing code were complete. so verifying and explaining is where the signal is
Great breakdown, Gregor. My takeaway for candidates: since big companies still run LeetCode-style rounds, the prep doesn't go away, but grinding 500 random problems is the wrong way to do it. Learning the ~20 underlying patterns lets you reason through problems you've never seen, which is also exactly what the open-ended alternatives test. It's the approach we've used at Design Gurus for years, and it holds up even better now.
Balancing a red-black tree was never relevant, or only in the rarest of cases. LeetCode is/was popular with large companies because 1. it selects for candidates who are ambitious and will do whatever it takes (interview prep etc) to meet an arbitrary goal (which makes it a predictor of success at many companies), and 2. it's more "objective", so easier to defend when the company gets sued.
The code review exercise is the one I'd pick too, with one tweak: let the candidate run the code. I recently gave the same buggy file to two AI reviewers. The one that only read it flagged the real cause as a medium issue, in the middle of the list. The one told to run it first marked it blocking, with the numbers in hand. Same model. Reading code is a skill; checking the output against the data is the habit that catches what runs green. I'd want to see which one a candidate reaches for.
Asking 'how did you know the output was correct?' might be the most important interview question of the AI era. It tests judgement, which is what we actually hire for. I wonder how many hiring managers are ready to assess that well.
I wrote about this last year and landed on LeetCode not being dead, just no longer the whole game. If everyone uses AI to learn, the interview becomes the one place a company can see your own reasoning. It's still the only filter that scales, but you also need projects and an understanding of systems.
The most convincing alternative here is not the open-ended question, it is the 200–500 line unfamiliar codebase round with 60 minutes to read it, find the issue, and fix it. That is the actual job. A candidate cannot pre-grind that, and an AI assistant does not remove the judgment call — it just moves it earlier, to deciding which fix is safe. My one worry: 60 minutes is tight enough that it may still reward speed over care. Would a longer round, or a follow-up “what would you change next week” question, separate the fast talkers from the careful builders?
Really glad to see open-ended questions as your first pick. In the interviews I run, the follow-up is where the signal is: "why this way and not the simpler one?" tells me more than any finished solution, with or without AI. I'd add one thing to Option 1: agree on what a strong answer looks like before the interview, otherwise open-ended can drift into each interviewer's favourite tangent and you lose the consistency that made leetcode attractive to big companies in the first place. Felix's point about dropping the coding round is interesting too. Do you think projects can replace the live conversation for juniors, or just give it a better starting point?
The split in this piece is not really big versus small companies. It is between interviews that test recall and interviews that test judgment on work the candidate has not seen before.
LeetCode survives at Google, Meta, Amazon and the rest, on your account, not because it predicts the job well — your own 200-problem grind, which got you a job unlike the work itself, is the counterexample — but because it is objective at volume. A company hiring thousands needs a score it can defend, even if the score measures preparation more than performance. Smaller companies can afford the alternatives precisely because they do not need that defense.
What the three alternatives share is that the candidate cannot walk in with a pre-solved answer. A code-review exercise, a product and technical deep dive, and pair programming with AI allowed all force the candidate to work with an unfamiliar system in real time. That is the real break from LeetCode, not the presence of AI. AI is allowed in the new formats, which means the test is no longer whether you can produce code, but whether you can steer it.
That creates a new way to game the interview, though. A fluent candidate can narrate trade-offs convincingly without having made them. Your own question list points to the antidote: “How did you know the output was correct?” and “What alternatives did you consider?” Those only work if the interviewer pushes past the first answer — change one constraint, and ask what breaks. Without that follow-up, a deep dive becomes a storytelling contest, and a pair-programming round becomes a prompt-fluency contest.
Matteo's experience in the comments, that candidates struggle more with the comprehension and refactor round, is the encouraging sign: struggle with unfamiliar code is much harder to rehearse than a template. My question is about the rubric for that round. If you score it on finding the right answer, preparation courses will soon grind that too. If you score it on what the candidate chooses not to change, and why, you are finally testing judgment. Which version are you seeing in practice?
We moved from a take-home assignment to a 90-minute code comprehension and live refactor round, allowing AI tools. We want to test how well a candidate reads and understands code, and which principles and best practices they apply when structuring a small project.
Surprisingly, candidates seem to struggle more with this new interview.
Thanks for sharing this! Yeah, you can’t gamify such type of an interview and it’s harder to prepare for it as well.
I don't think
LeetCode isn't going away because most interviews are bad at measuring everything else and comfortable measuring code recall. The real question is what replaces the pattern matching part when AI handles it, I think it is explaining your tradeoffs on a messy real system. Are you seeing companies shift to live design or incident-style questions yet?
System design interview is more important than ever, that's my feeling talking with different folks. Additionally, I do see a "Debugging an existing application" type of exercise. Where the candidate is given a working application with a hidden bug, and is asked to investigate it on the interview.
agree, I think the trend will be less leetcode style, more AI assisted. At least ten companies already run a coding round with an AI assistant in the room, up from one (Canva) in June 2025.
the debugging exercise you mention fits that well. In practice data from Nyrion (I run it, a practice platform for these rounds) people accepted AI’s code 93% of the time, and fewer than 1 in 10 answers explaining their own passing code were complete. so verifying and explaining is where the signal is
Great breakdown, Gregor. My takeaway for candidates: since big companies still run LeetCode-style rounds, the prep doesn't go away, but grinding 500 random problems is the wrong way to do it. Learning the ~20 underlying patterns lets you reason through problems you've never seen, which is also exactly what the open-ended alternatives test. It's the approach we've used at Design Gurus for years, and it holds up even better now.
AI shifts interviews toward collaboration, pair programming and projects. Will take home projects become the standard?
Balancing a red-black tree was never relevant, or only in the rarest of cases. LeetCode is/was popular with large companies because 1. it selects for candidates who are ambitious and will do whatever it takes (interview prep etc) to meet an arbitrary goal (which makes it a predictor of success at many companies), and 2. it's more "objective", so easier to defend when the company gets sued.
The code review exercise is the one I'd pick too, with one tweak: let the candidate run the code. I recently gave the same buggy file to two AI reviewers. The one that only read it flagged the real cause as a medium issue, in the middle of the list. The one told to run it first marked it blocking, with the numbers in hand. Same model. Reading code is a skill; checking the output against the data is the habit that catches what runs green. I'd want to see which one a candidate reaches for.
Asking 'how did you know the output was correct?' might be the most important interview question of the AI era. It tests judgement, which is what we actually hire for. I wonder how many hiring managers are ready to assess that well.
I wrote about this last year and landed on LeetCode not being dead, just no longer the whole game. If everyone uses AI to learn, the interview becomes the one place a company can see your own reasoning. It's still the only filter that scales, but you also need projects and an understanding of systems.
The most convincing alternative here is not the open-ended question, it is the 200–500 line unfamiliar codebase round with 60 minutes to read it, find the issue, and fix it. That is the actual job. A candidate cannot pre-grind that, and an AI assistant does not remove the judgment call — it just moves it earlier, to deciding which fix is safe. My one worry: 60 minutes is tight enough that it may still reward speed over care. Would a longer round, or a follow-up “what would you change next week” question, separate the fast talkers from the careful builders?
Really glad to see open-ended questions as your first pick. In the interviews I run, the follow-up is where the signal is: "why this way and not the simpler one?" tells me more than any finished solution, with or without AI. I'd add one thing to Option 1: agree on what a strong answer looks like before the interview, otherwise open-ended can drift into each interviewer's favourite tangent and you lose the consistency that made leetcode attractive to big companies in the first place. Felix's point about dropping the coding round is interesting too. Do you think projects can replace the live conversation for juniors, or just give it a better starting point?
The split in this piece is not really big versus small companies. It is between interviews that test recall and interviews that test judgment on work the candidate has not seen before.
LeetCode survives at Google, Meta, Amazon and the rest, on your account, not because it predicts the job well — your own 200-problem grind, which got you a job unlike the work itself, is the counterexample — but because it is objective at volume. A company hiring thousands needs a score it can defend, even if the score measures preparation more than performance. Smaller companies can afford the alternatives precisely because they do not need that defense.
What the three alternatives share is that the candidate cannot walk in with a pre-solved answer. A code-review exercise, a product and technical deep dive, and pair programming with AI allowed all force the candidate to work with an unfamiliar system in real time. That is the real break from LeetCode, not the presence of AI. AI is allowed in the new formats, which means the test is no longer whether you can produce code, but whether you can steer it.
That creates a new way to game the interview, though. A fluent candidate can narrate trade-offs convincingly without having made them. Your own question list points to the antidote: “How did you know the output was correct?” and “What alternatives did you consider?” Those only work if the interviewer pushes past the first answer — change one constraint, and ask what breaks. Without that follow-up, a deep dive becomes a storytelling contest, and a pair-programming round becomes a prompt-fluency contest.
Matteo's experience in the comments, that candidates struggle more with the comprehension and refactor round, is the encouraging sign: struggle with unfamiliar code is much harder to rehearse than a template. My question is about the rubric for that round. If you score it on finding the right answer, preparation courses will soon grind that too. If you score it on what the candidate chooses not to change, and why, you are finally testing judgment. Which version are you seeing in practice?