Our 4 Interviewers Agreed on Nothing and We Hired the Wrong Engineer Twice: the 7-Step Debrief That Fixed It

A small group reviewing notes around a table, illustrating a structured developer interview debrief and hiring committee decision in a Singapore engineering team
William

William

Talent Sourcing Expert · November 12, 2026 · 10 min read

TL;DR

  • •The problem: we ran good interviews and made bad decisions. 4 interviewers, four overlapping opinions about general ability, and a debrief that was really a negotiation between whoever spoke first and whoever was most senior.
  • •The method: 7 steps. The one that did most of the work is step 2, written evidence submitted before anyone speaks or sees another submission. It is also the step every team skips.
  • •The honest part: both wrong hires were people I had argued for. Step 7 exists because the only way to know whether a debrief works is to check it against what the hire actually did 6 months later.

The second wrong hire is the one that forced this. Capable engineer, wrong role, gone in five months, and when I went back through the notes from the debrief I found something uncomfortable: nobody had actually said the candidate was strong. What had happened was that I had said it first, nobody had evidence to contradict me, and the absence of objection had been recorded as agreement. The interviews were fine. The decision was the broken part.

The Failure Is in the Decision, Not the Interview

Most engineering teams in Singapore have spent real effort on their interviews and almost none on the thirty minutes afterwards where the decision is made. That is the wrong allocation, because the debrief is where good evidence gets thrown away.

Three things go wrong, reliably. Anchoring: the first confident speaker sets the frame and everything after it is a reaction. Seniority weighting: the most senior person in the room is right more often than chance but far less often than their influence implies. Absence treated as assent: an interviewer with no evidence on a dimension says nothing, and silence gets counted as a yes. That last one is what sank my second bad hire.

None of those are fixed by better interview questions. They are fixed by changing the order of operations, which is what the seven steps below do. Five of them cost nothing.

Step 1: Assign Non-Overlapping Coverage Before Anyone Interviews

Before the first interview is scheduled, write down which interviewer covers which area, and make sure no two people cover the same one.

This is the structural fix for the four-opinions problem. If four people each assess general engineering ability, you finish with four views of one dimension and no evidence at all about the other three that matter. Assign them: one takes systems reasoning, one takes code quality in their actual working language, one takes collaboration and how they handle being wrong, one takes the specific domain or stack risk for this role.

Write it in the scheduling invitation, not in someone head. An interviewer who knows they own collaboration asks about collaboration. An interviewer who thinks they own everything asks whatever occurs to them and reports a feeling.

Our expert take

The coverage map has a second use that is almost more valuable than the first. When a hire does not work out, you can look back and see which dimension nobody actually assessed. In both of our bad hires the answer was the same: nobody owned the question of whether this person wanted the job that existed, as opposed to the job in the advert. That is now an assigned area with a named owner, and it has changed two decisions since.

Step 2: Collect Written Evidence Before Anyone Speaks

If you adopt one step from this piece, adopt this one. Every interviewer submits written notes and a provisional score before the debrief, and before seeing anyone else submission.

This single rule removes anchoring and seniority weighting at once, because every position on the record was formed independently. It also converts “I liked them” into something examinable, since a written note that contains no evidence is visibly a note that contains no evidence, in a way that a confident sentence in a meeting is not.

The submission should contain what the candidate actually said or did, then the score, in that order. Notes written the other way around tend to be justifications. A deadline matters too: within four hours of the interview, while recall is real rather than reconstructed.

Expect resistance, and expect it to be about time. The honest counter is that it takes about ten minutes, and the alternative costs a quarter of someone salary when it goes wrong.

Where the evidence goesSame four interviews, two ways of combining themUnstructured debriefSenior speaks first: frame setInterviewer 2 adjusts toward the frameInterviewer 3 has no evidence, stays quietSilence recorded as agreementDecision: 1 opinion, 4 signaturesWritten submission first4 independent notes, unseenEvidence before score, per assigned areaGaps are visible as gapsDisagreement surfaces instead of dissolvingDecision: 4 opinions, 1 deciderBoth columns describe our own debriefs before and after March 2026, not an industry benchmark.The failure in the left column is the fourth row: an interviewer with nothing to report is not an interviewer who agrees.Note that the right column does not reduce disagreement. It makes it visible early, which is the point.

Step 3: Score Against the Scorecard, Not Against Each Other

The question a score answers is narrow: did this candidate meet the bar for the area I was assigned. Not whether they were better than the person interviewed yesterday, and not whether they are as good as the interviewer was at the same stage.

Comparative scoring produces drift. A strong week raises the bar, a thin pipeline lowers it, and six months later nobody can say what the standard is. The tell is an interviewer who says a candidate was “the best we have seen this month”, which is information about the month rather than about the candidate.

If you do not have a scorecard yet, that is the prerequisite rather than this article: we cover the construction in our method for evaluating full-stack developers.

Step 4: Run the Debrief in a Fixed Order, Weakest Signal First

Open with the area where the evidence is thinnest, not with the person who has the strongest view.

Two reasons. It puts attention where the risk is, since the dimension nobody has evidence on is the one most likely to produce a surprise in month three. And it defuses the seniority effect without anyone having to be told to speak less, because the order is set by the evidence rather than by the room.

The decider speaks last. Always. A decider who opens the debrief has, in practice, ended it.

Want us to look at a debrief that went wrong

Send us the notes from a hire you regret, with names removed. We will show you which of the seven steps was missing and which dimension nobody actually assessed.

Get started

Step 5: Force a Decision, With Four Allowed Outcomes

Exactly four outcomes are permitted, and the debrief does not end until one is chosen:

  1. Hire, at a stated level.
  2. No hire.
  3. Hire at a different level than advertised, which is the outcome teams forget exists and the one that rescues the most candidates.
  4. One more interview, with a named interviewer and a written question it must answer.

What is banned is the fifth outcome: the vague maybe. “Let us keep them warm” is how a pipeline silts up with unresolved candidates who eventually accept somewhere else and tell their network you were slow. If option four is chosen without a specific question attached, it is really the banned option wearing a disguise.

Do not average scores. A mean of a strong yes and a strong no is a decision nobody in the room holds, and it is the mechanism by which teams hire candidates no individual interviewer would have defended.

Four doors out of the room, and one that is bricked upThe debrief does not end until one of the first four is chosenHireat a stated levellevel is part of thedecision, not after itNo hirewith a written reasonand the counterfactualfrom step 6Different levelthe forgotten outcomerescues more candidatesthan any otherOne more interviewnamed interviewer anda written question itmust answerBanned: the vague maybe"Let us keep them warm" is how candidates are lost to silenceOption four without a specific written question attached is the banned fifth option wearing a disguise.Never average the scores: a mean of a strong yes and a strong no is a decision nobody in the room actually holds.

Step 6: Record the Reason and the Counterfactual

Two sentences in the record. Why the decision went this way, and what would have changed it.

The counterfactual is the part that earns its keep. It makes the record useful in the three situations that actually arise: the candidate reapplies in a year and you can tell them something specific, the decision turns out to be wrong and you can see which assumption failed, or a similar candidate appears and you can check whether you are being consistent.

This also matters for fairness in a market as international as Singapore. A written reason is reviewable. A remembered impression is not, and impressions are where inconsistency hides, usually without anyone intending it.

Step 7: Audit Decisions Against Performance at Six Months

Twice a year, put the debrief records next to how those hires actually performed. You are looking for three patterns:

  • Which interviewer predicted well. Some people are consistently right and quiet. Those are your deciders, and they are frequently not the most senior person available.
  • Which signals predicted nothing. Every team carries at least one interview question it is attached to and which has never once correlated with performance. Find it and drop it.
  • Which dimension keeps being the surprise. If month-three surprises cluster in one area, that area is under-covered in step 1 regardless of what the map says.

Keep it small. Two hours, twice a year, on the hires you have six months of evidence for. The exercise fails when it becomes a performance review of the interviewers rather than of the process.

StepCostThe failure it prevents
1. Coverage map15 min per roleFour opinions on one dimension, none on three others
2. Written submission first10 min per interviewerAnchoring and seniority weighting
3. Scorecard not comparisonZeroBar drifting with pipeline quality
4. Fixed order, weakest firstZeroAttention going to the loudest view
5. Four outcomes, no maybeZeroCandidates lost to silence
6. Reason and counterfactual2 minUnreviewable, inconsistent decisions
7. Six-month audit2 hours, twice a yearA process that never learns

The 4 Things This Will Not Fix

  1. A bad role definition. If nobody agrees what the person will do, a structured debrief produces a well-documented argument about an undefined job. Fix the role first; the debrief will make the confusion visible, which is useful but not the same as solving it.
  2. Interviewers who do not know what good looks like. Structure organises judgement, it does not create it. A panel that has never worked with a strong engineer in a given area will produce confident, well-formatted notes about the wrong things.
  3. A pipeline with nobody in it. Decision quality is irrelevant at the top of an empty funnel. If you are choosing between two candidates because two is all you have, the work is upstream in sourcing.
  4. Compensation that is not competitive. You can run an excellent process and lose every candidate at offer. If that is your pattern, the debrief is not the problem and tightening it further will not help.

For the stages around this one, see our methods for running remote technical interviews and reference checks on Singapore developer hires, and for the build-versus-hire question our guide to custom software development in Singapore.

Frequently Asked Questions

How many people should be in a developer hiring debrief?
Three or four interviewers plus a named decider is enough for almost every engineering role, and more is usually worse. Each additional interviewer adds scheduling delay, dilutes accountability and increases the chance that one person will have a weak objection that stalls the decision. The useful constraint is coverage rather than headcount: if your four interviewers cover four genuinely different areas, four is plenty. If seven people all assess general problem solving, seven is a committee with one data point. For a senior or lead role you may want a fifth voice for the cross-team dimension, but that person should have an assigned area rather than a general impression.
What if the interviewers genuinely disagree?
Disagreement is the normal and healthy case, and the process exists for it. First check whether it is a disagreement about evidence or about the bar. Two interviewers describing the same answer differently is an evidence problem and usually resolves by reading the written notes from step two. Two interviewers who agree on what the candidate said but disagree on whether it is good enough is a bar problem, and that is the decider job to settle rather than the room. What you must not do is average the scores: a mean of a strong yes and a strong no is a decision nobody in the room actually holds, and it is how teams hire candidates no interviewer would have defended.
Does this slow hiring down in a competitive market like Singapore?
The debrief itself takes thirty to forty minutes, so the direct cost is small. The process adds a day or two at the front in step one, and removes considerably more at the back by eliminating the undecided candidate who sits in the pipeline for three weeks. In a competitive market the thing that loses you candidates is not a structured decision, it is silence after the final round. A process that reliably produces an answer within forty-eight hours of the last interview is a competitive advantage rather than a tax. If your debrief is adding a week, the problem is scheduling it, which step four helps with because the debrief no longer needs everyone present to be useful.
How does this work when part of the panel is in another timezone?
Better than a conversation does, which is part of why step two exists. Written submission before the debrief means a remote interviewer contribution carries the same weight as somebody in the room, rather than depending on whether they could attend a call at an awkward hour. In a distributed panel I would run the whole first pass asynchronously: everyone submits, the decider reads, and a live call happens only where there is genuine disagreement about the bar. That turns a scheduling problem into a reading problem, and it produces a better record, because written reasoning tends to be more careful than spoken reasoning and it is available later for the six-month audit in step seven.

Start with the coverage map, not the meeting

We help Singapore employers tighten the decision end of hiring, then shortlist pre-vetted developers who can be assessed against a real scorecard rather than a general impression.

Get started