I have run two onsites in the last three years where I bombed round 2 of 5. The first time I got the offer. The second time I did not. The two loops were not identical, but they were close enough that the difference is mostly about what I did in the rounds that came after the bad one. I am going to walk through both loops in parallel, because the comparison is the entire point.

The two loops

Loop A was at a mid-size FAANG-adjacent company (about 800 engineers, B2B platform, senior engineer role). Five 45 minute rounds, all in one day, virtual. Round 2 was a coding round and it went badly: I solved the optimal version with 4 minutes left, bug-ridden, never wrote a test. I left round 2 thinking the loop was probably over.

Loop B was at a Series D infrastructure company (about 600 engineers, similar role level). Same shape: five 45 minute rounds, virtual, one day. Round 2 was a system design round and it also went badly: I anchored on the wrong abstraction (cache layer first, when the prompt was really about write throughput), recovered halfway through but never got out of the hole.

Both loops had the same break in the same round. The first one ended in an offer. The second one ended in a rejection. What was different was the next 30 minutes after the bad round in each loop.

Loop A: what I did between rounds

I had a 30 minute break between round 2 and round 3. Here is what I did:

minute 0-3   walked away from the desk, drank water, did not look at slack
minute 3-10  named the failure mode out loud to myself in one sentence
minute 10-15 wrote down two things I would not do in the next round
minute 15-25 reviewed my notes for round 3 (system design) and forced myself to
             ask the structuring question first, no matter what the prompt was
minute 25-30 deep breaths, water, back to the desk 90 seconds early

The key move was minute 3 to 10. Naming the failure mode out loud was the trick. The sentence I came up with was: "I started typing too early because I was trying to look fast". That sentence became the anchor for the next three rounds. In round 3, when the system design prompt came up, I caught myself wanting to start drawing in the first minute, made myself ask the clarifying question instead, and the round opened up. Round 4 (a behavioral) and round 5 (a take-home review) were strong. The hiring manager told me later that the debrief had been split until round 4, where I had told a story about a project I had killed; that story had moved one of the holdouts.

Loop B: what I did between rounds

I had a 30 minute break between round 2 and round 3 in this loop too. Here is what I actually did:

minute 0-15  ruminated on round 2, replayed the bad moments, googled the right
             answer, felt worse
minute 15-22 ate something, did not really taste it
minute 22-30 reviewed notes for round 3 but only half-attentively

I walked into round 3 carrying round 2. I have a recording of myself from that round (the company shared it after asking permission, as part of their feedback process), and I can hear the carry-over in my voice. I was talking faster, hedging more, jumping to solutions before the interviewer had finished the prompt. The round was a coding round; I solved it but the conversation was off. Rounds 4 and 5 stabilized but the deficit from rounds 2 and 3 was too deep to recover.

The rejection came two business days later. The recruiter told me, with the candor that good recruiters use, that the debrief had been a unanimous no-hire by the time round 3 ended. Rounds 4 and 5 had not changed anyone's vote.

What the recording showed, when I went back and listened to it three days later, was specific. The first symptom was pace: I was answering inside two seconds of the interviewer finishing the prompt, where in a calm round I would take five to seven seconds and structure the answer first. The second symptom was hedging language: I used the words "I think" and "probably" eleven times in a 45-minute round where I would normally use them twice. The third symptom was the ratio of clarifying questions to solution attempts. In a calm round I ask three to five clarifying questions before drawing anything; in this round I asked one and started typing. The fourth and worst symptom was that I had stopped asking the interviewer questions back. The whole round read as the candidate trying to prove competence rather than collaborating on a problem. None of those symptoms were visible to me in the moment. They were obvious on the recording.

The single thing that was different

The gap between the two loops was the 30 minutes between rounds. Specifically: in loop A I named the failure mode out loud and used it as an anchor for the next round. In loop B I let the failure mode rumble in my head silently and let it leak into the next round.

This is the actual mechanic of mid-loop recovery, and it is smaller than most advice posts make it out to be. The advice I had read before loop A talked about "bouncing back" and "compartmentalizing" and "resetting". Those are descriptions of the outcome, not instructions for how to get there. The instruction that worked for me, derived in retrospect, was specific:

the recovery move
  1. step away from the desk for at least 3 minutes
  2. say the failure mode out loud in one sentence (out loud, not in your head)
  3. write down what you will not do in the next round
  4. spend the rest of the break preparing the next round, not relitigating
     the last one

The out-loud part is what makes step 2 work. Saying it in your head lets your head keep saying it. Saying it out loud, with your own voice, ends the loop. I have used this in three loops since and it has never failed to land me in the next round in a different cognitive state from the one I left the bad round in.

What about not bombing round 2 in the first place

This is the right question to ask, and the answer is that the bombing is partly bad luck. In loop A the round 2 problem was a graph variant I had not seen and I underestimated the time to write the optimal version. In loop B the round 2 prompt was a system that was off-shape from the systems I had drilled on. You can drive the bombing rate down with prep, but you cannot drive it to zero, and the assumption that you can is what makes the bombing feel catastrophic when it happens.

The more useful frame: in any 5 round loop, treat one bombed round as the expected baseline. Two bombed rounds is the danger zone. The recovery move is the difference between one bomb and two.

The four-step recovery sticky note

The loop A offer was the right job and I took it. The loop B rejection was clean and I recovered from it (got an offer at a different company two months later). Two things I would do differently next time.

The first is to plan the recovery move before the loop, not in the moment. Before any onsite I now write the four-step recovery protocol on a sticky note and put it on my monitor where I will see it during the break. The recovery move is muscle memory you have to install in advance; you cannot install it while panicking.

The second is to remind myself, before the loop starts, that one bad round is the baseline. The candidate who walks into round 1 expecting all 5 rounds to be perfect is the candidate most likely to spiral when the bad one happens. Lowering the bar to "4 out of 5 strong, 1 acceptable" before the loop starts is the single most effective piece of mental prep I now do.