Sarthak Garg

Critique without crushing

Say hard things in a way the receiver can act on.

·10 min read·

Critique means telling someone what is wrong with a piece of work they made, such as a design doc or a pull request, in a way they can use to make the work better. Two well-known frameworks describe how to do it well: Kim Scott's Radical Candor defines good feedback as caring personally while challenging directly, and Scott calls the most common failure ruinous empathy, which is when the manager likes the person too much to say the hard thing and the problem ships unfixed. Alongside it, the Situation, Behaviour, Impact model from the Center for Creative Leadership asks the giver to describe what the person did in terms a camera could record, tied to a specific moment and what it caused, so the critique stays on the work and does not drift into the person's character.

I use both, and my notes got better, but they did not stop me from sending notes that were candid and kindly meant and that the receiver read as a verdict on them. When I asked afterwards what they had taken from one of those notes, they had read my words correctly and filled in everything I had left out with the worst available guess, which is a sensible strategy when your manager has just told you something is wrong.

A hard note leaves the receiver with questions that the note itself rarely answers:

  • How serious is the problem?
  • Why is this person raising it?
  • Am I being told, or advised?
  • What exactly is wrong?

When the receiver has to answer those alone, they tend to pick the more threatening answer, and a note that was meant to help them crushes them instead. Each of those questions can be answered in the note itself, in a sentence.

Read the note as they will

When you write a hard note, you know how much you respect the person and what you would accept as a fix, and the receiver gets none of that, only the words. A 2022 study of code review by Sanuri Gunawardena and colleagues at the University of Auckland and Google found that the people who give notes and the people who receive them see the same notes differently. The researchers asked engineers about destructive criticism, which they defined as negative feedback that is both nonspecific and inconsiderate. More than half of the respondents had received nonspecific negative feedback in the past year, and as a group they reported receiving destructive criticism more often than they reported giving it. Both figures can be true only if people fail to notice some of the destructive criticism they give, so your sense that a note was clear and kind is a poor guide to how it will read.

Before the note goes out, test it on the two things the study measured: whether it names the part of the work and what is wrong with it, and whether it stays on the work without saying anything about the author's ability. The basketball coach John Wooden, observed for a season by the psychologists Roland Tharp and Ronald Gallimore, corrected players in a few seconds by showing them their own version between two demonstrations of the right one, and a written note passes both tests most easily when it does the same and puts the better version next to the current one.

A reviewer reads a design doc and writes "this section is confusing" next to the retry logic. To the reviewer it is a quick flag, but the author learns nothing about where the confusion starts and reads it as a comment on their writing. The rewritten note says: "I lost track of which service owns the retry here. I read it twice and still could not tell whether the gateway retries or the worker does."

After the note has landed, either once the author has replied or at the end of the conversation, ask what they plan to change first, in their own words. If they name a small fix when you meant a structural one, you have found the gap while it is still cheap to close.

Open with the reason

In a 2014 study led by David Yeager and Geoffrey Cohen, seventh-grade students received teacher comments on their essays, and half of them also got one extra line saying the teacher gave these comments because they held high standards and believed the student could reach them. Students who got that line revised more often, and the effect was largest among students who had reason to distrust their school. When a critique arrives without a stated reason, the receiver supplies one, and a common one is that the giver thinks poorly of them.

Put the reason in the first sentence, because a receiver who meets the problem first reads everything after it defensively. The old praise sandwich wraps the critique in compliments that have nothing to do with it, and receivers learn to skip them, but a stated reason tells the receiver that the note is meant to be acted on, so they take it more seriously and nothing in it has been softened.

A lead reviewing a newer engineer's first design doc opens with: "This is going to the architecture review next week, so I am holding it to the bar that group will use, and I think you can get it there in one more pass." The notes that follow are as hard as they would have been without that sentence, and the engineer reads them as a route to the review.

The sentence has to be true, because a manager who pastes it onto every note, or who does not believe the person can meet the bar, teaches the team to ignore it. It also depends on trust that existed before the note did, because a team that does not trust its manager hears the same line as flattery, and Trust as a measurable thing covers how to tell which kind of team you have.

Ask before you assert

A question costs the reviewer one more round of replies, while an opinion that turns out to be wrong about a choice the author made for a reason they did not write down costs more, because the reviewer is now wrong in writing and the author has to correct them. So when you do not know why the author made a choice, ask before you give an opinion. The choreographer Liz Lerman built her Critical Response Process, a method for feedback on dance and other creative work, around this order, with neutral questions before any opinion. A neutral question such as "What load did you size this for?" has no opinion hidden inside it, whereas "Did you consider that this will not scale?" is an opinion with a question mark on the end, and the receiver hears both the opinion and a test they have apparently failed.

A reviewer about to write that a message queue is the wrong tool asks first what failure the author was designing around, and learns about a constraint from another team's service that the doc never mentioned. The critique changes from "wrong tool" to "this constraint needs to be in the doc, because the next reader will ask the same question."

If the objection still holds after the answer, say it plainly in your next reply.

Send the structural note alone

Book editors send their large notes, about structure or a character who does not work, in a separate editorial letter, and the line edits go into the manuscript. Authors say they skip straight past the praise in those letters, which seems fair. When the big note arrives on its own, the author cannot mistake it for one more small fix.

A design doc comes back with forty inline comments. The one that says the data model will not support the second use case sits at number twenty-three, between two notes about naming. The author fixes the naming notes first, because they are quick, and reaches the data model comment with half their energy spent on sentences the real problem will force them to rewrite.

When the structure is in question, send that note first and by itself, ideally as a conversation rather than a comment, and hold the line notes until the structure is settled. Polishing a section that may be rewritten wastes the author's effort and tells them the small things count as much as the large one.

Say what is optional and what is not

Inside a code review thread, a team can label how much each comment matters, and Code review culture covers those labels. Outside a review tool, in a design discussion or a conversation after a demo, people signal how much a note matters with small hedging words, and the receiver has to work out what those words mean from their own habits. Erin Meyer, who studies how working cultures differ, describes how some cultures soften criticism with words such as "maybe" and "a little" while others strengthen it with words such as "totally," so the same sentence can read as optional to one receiver and as harsh to another.

On a team spread across three countries, a reviewer writes "might be worth reconsidering the caching approach here," meaning that it has to change before launch. The author, used to a more direct style, reads a suggestion and ships, and neither of them did anything careless, because the severity was never written down anywhere except in the reviewer's head.

Airline crews are trained to raise a concern with fixed words that run from "concerned" up to "safety issue," so that a first officer and a captain both know how serious it is without reading anything into tone. An engineering team needs only a few plain phrases, used the same way every time:

  • Take it or leave it: my preference, and you decide.
  • I would want this changed: I think it matters, and I expect a reply if you disagree.
  • I cannot approve this as it is: it has to change before it ships.

Lerman's process also lets the maker decline an opinion, and Pixar's Braintrust, the senior filmmakers who review films in progress, has no authority over the director, so in both cases the person who made the work decides what to take. A manager does have authority, so say plainly in the same note who decides: which points the owner may reject and which are a bar you are holding, with the reason for each.

A manager reviewing a launch plan writes: "Two of these are mine to hold, the rollback path and the data retention window, because we have promised both to customers. The rest are my opinion, and you know this system better than I do, so take what helps."

A manager who calls everything a suggestion while expecting it to be done is not fooling anyone, and teaches the team to read every suggestion as an order.

Aim at the judgement in generated work

When much of a change was written by a model or an agent, a note about a bad line has no obvious person to land on, and a question about who wrote it lands squarely on the author's sense of their own competence. The author can act on a note about the decisions that stayed theirs: what they asked the tool for, and what they checked before they kept the result. (Reading code in the era of generated code covers how to read the change itself.)

A change includes a generated retry helper that silently swallows errors. The useful note says: "The error path is the part we cannot let a generator decide for us, so before this comes back, please read this helper line by line and decide what it should do when the call fails." I do not have a settled view on how much of a generated change an author should have read before asking for review, and I suspect the answer is different for an error path than for a test fixture.

When not to

Marcus Buckingham and Ashley Goodall argued in the Harvard Business Review in 2019 that telling people where they fall short tends to inhibit their learning. Their argument is about feedback on people and their habits, and critique of one piece of work, with the reason and the size stated, is a narrower thing that a team still needs whenever the work is going somewhere other people depend on.

A note is the wrong tool when the problem is a pattern in the person rather than in this piece of work, because a single note will not fix a pattern and the repair loop handles it better. And when the work is sound and only differs from how you would have done it, mark the note as take it or leave it, or keep it to yourself.