The symptom is familiar: the same class of mistake keeps reaching production. A wrong environment variable, a forgotten migration, a config value copied from the wrong file. Each time, the team fixes it, writes a note and moves on, and a few weeks later it happens again. Skill is seldom the culprit. More often, the work is being checked only by the person who produced it. Adding a second pair of eyes at the right moments is the practical remedy, and this article walks through the problem, the causes and the fix in the order you would meet them on the job.
The symptom: errors that were “obvious in hindsight”
Listen to how incident write-ups describe these failures. The phrases “should have been caught,” “it was a small typo” and “we all assumed” appear again and again. Those phrases point to errors that a fresh reader would have noticed in seconds, but which the author, deep in context, could not see. When your review notes keep saying the same thing, treat that as data.
Diagnosis one: solo checking has a blind spot
People who build a change verify what they intended. They test the path they imagined and skim the parts they trust. That is human, not negligent. The cure is not “be more careful”; it is to place a different viewpoint on the work before it goes live. Think of a proofreader for a printed document: the writer cannot be the proofreader, no matter how conscientious.
Fix: require a named partner for any change that reaches a live system. The partner reads the change, not a summary, and confirms the specific things you list.
Diagnosis two: the checks happen too late
Many teams do have review, but it sits at the end, after the author has invested days. By then a reviewer who finds a design flaw faces social pressure to wave it through. A better arrangement catches problems earlier, when a quick conversation is still cheap.
Fix: hold a ten-minute pairing session at the start or midpoint of any non-trivial task. Show your approach and ask, “What am I missing?” The same questions that expose design gaps early are well covered in this practical guide for operations teams planning their next steps with intelligent automation, which shows how early planning conversations prevent late surprises.
Diagnosis three: incidents are handled by one tired person
During an outage, a single engineer often investigates alone while everyone else waits. Fatigue narrows attention and invites tunnel vision, and the habit of checking again and again goes unchallenged. The longer the incident drags on, the more likely it is that the responder repeats the same checks or misreads a graph.
Fix: pair every serious incident. One person drives; the other narrates what has been tried, writes down findings and challenges assumptions. The partner also notices when the responder has gone twenty minutes without trying anything new. For further reading on structuring response work, see this walkthrough of faster incident resolution with automated assistance, and apply its ideas with a human partner alongside any tooling.
Diagnosis four: nobody is rewarded for catching mistakes
If reviewers are seen as obstacles and authors as heroes who ship quickly, review will be cursory. People respond to what is noticed.
Fix: publicly credit catches. In a retrospective, say “Priya’s review caught the wrong region setting.” Keep the tone appreciative and never use review findings in performance rankings against authors.
Diagnosis five: paperwork and logistics get no review at all
Code gets attention, but surrounding tasks do not: access requests, vendor forms, travel bookings for on-site work, purchase orders. These are full of names, dates and amounts, and errors in them waste days.
The principle applies far outside engineering. A medical tourism website like KneeHospitals, where hospitals and doctors are presented per country and visitors fill in quote-request or appointment forms, illustrates this kind of form-based process, and anyone submitting such forms for themselves or a relative does well to ask someone else to read them first. The check concerns accuracy of the entries only and makes no medical judgement.
Fix: extend your buddy rule to any document with consequences, and use a short checklist: names, dates, amounts, attachments, destination.
A case from the field
A team kept shipping releases with a stale feature flag left switched on in one region. Each time, the author had checked the flag in the primary region and moved on. After the third occurrence they added a rule: the reviewer must open the settings for every region and read the values aloud to the author on a call. The first time they tried it, the reviewer spotted a stale flag within a minute. Over the next quarter the “wrong region” entry disappeared from their incident notes. Nothing about the engineers changed. Only the viewpoint did.
Cost and effort: a realistic estimate
Teams worry about the time. Count it honestly. A fifteen-minute review on each of, say, six changes in a week is ninety minutes of reviewer time and about the same again of author time spent responding. Compare that to even one avoidable incident, which normally consumes several people for hours plus the cleanup afterwards. The sums favour review comfortably, and the gap widens if the incident affects customers.
Prescription: a minimal buddy system
- Define the trigger. Anything that touches production, money or a deadline needs a second reader.
- Name the partner in advance. A rota avoids the “who is free?” delay.
- Give the reviewer a task list. Three to five concrete checks beat “have a look.”
- Cap the effort. Fifteen minutes per change keeps the practice sustainable.
- Log what was found. Patterns in the log tell you what to automate or teach next.
How to confirm the diagnosis was right
Count repeat errors for a month, then introduce the buddy rule and count again. You are looking for fewer “should have been caught” entries, not for zero. Also ask the team how the process feels. If reviewers report feeling rushed, widen the rota. If authors feel judged, revisit the tone of comments.
Fitting the rule into a busy week
Reserve a recurring half hour, perhaps just after the morning stand-up, as the default review window. Partners know when to expect requests, and authors learn to prepare changes in time for it. Urgent fixes can still jump the queue, but they get the shortened three-check version rather than nothing, which keeps the rule alive when pressure is highest.
What not to do
- Do not make one senior person the reviewer for everything. They become a bottleneck and run out of attention.
- Do not accept “looks good” as a review. Ask what the reviewer verified.
- Do not skip the buddy rule during emergencies. Urgent changes are the riskiest ones.
The fix for repeat mistakes is rarely a new tool. It is a habit: before something important goes out, someone who did not make it takes a short, structured look.

