A language model writes kinder, clearer, better-organized feedback than I do. I’ve stopped arguing about that, because it’s true and because it’s the wrong thing to argue about.
The argument people build on top of it is the one I want to take apart. It goes: the technical half of engineering leadership is getting automated, the soft half is next, and anyone whose value is “good with people” should be nervous, because the machine is already better at the sentence. That last clause is right. The conclusion is wrong, and it’s wrong in a specific way. The hard part of this job was never the wording.
What the model is genuinely good at#
Give it a draft of a hard message and it will find the accusation buried in the third paragraph, the passive construction hiding who decided what, the sentence that reads as a question but functions as a verdict. It will hand back something more generous than what I wrote while I was annoyed. That’s real, and I use it for exactly this.
So this isn’t a piece about how the machine can’t write. It writes fine. It’s a piece about everything that has to happen before there is anything to write.
Three things that happen before the message exists#
Noticing that there’s something to address at all. Nobody files a ticket that says a teammate has gone quiet in retro, or that the engineer who used to leave long review comments now leaves “lgtm,” or that someone’s estimates have gotten pessimistic in a way that has nothing to do with the estimates. There’s no artifact. The model has no context window that contains it, because the information never became text. Somebody has to notice, and noticing is the scarce thing.
Not fixing it immediately. The righting reflex is Motivational Interviewing’s name for the urge to solve what looks broken the second you spot it. I once got three sentences into a solution before a teammate had finished describing the setup. The model has the same reflex, permanently and structurally: you ask it a question and it answers. It has no mechanism for deciding that the useful move right now is one more question and no advice. That call is judgment about a specific person on a specific day, and it’s the difference between a 1:1 that helps and one that performs helping.
Using their words instead of yours. In a 1:1 a teammate once told me they felt “stuck,” and I nearly replied “stuck like you need a break?” That would have handed them my metaphor and quietly made the problem mine to solve. Clean language is the discipline of handing their exact word back instead. A model asked to be empathetic will reach for the fluent paraphrase every time, because fluency is what it optimizes. Fluent paraphrase is the failure mode here, not the goal.
The specificity is the whole thing#
I once left a comment on a pull request that said the retry logic was good work, and named the edge case that made it easy to get wrong. The response told me it landed harder than the approval did. Transactional Analysis would call that a stroke, and TA’s claim is that we’re hungry for recognition roughly the way we’re hungry for food. I wrote that one down because of the size of the gap between “nice work” and naming the thing that was actually hard.
A model can generate that second sentence. It’ll do a nice job. But it can only do it if you tell it which part was hard, and knowing which part was hard for this engineer, as opposed to hard in general, is not a writing problem. It’s the entire skill, and it comes from having paid attention to that person over months. The sentence is the last five percent.
The failure mode gets worse, not better#
Here’s what actually concerns me, and it isn’t the machine replacing anybody.
Clumsy words that show you noticed something real beat perfect words aimed at the wrong thing. Everyone has received the second kind. It reads as fluent, considerate, and slightly off, and the effect is worse than saying nothing, because now the person knows you’ve spent effort on a version of them that isn’t quite who they are. That message used to be hard to produce. Now it takes eight seconds.
This is the thing I keep coming back to: AI amplifies whatever you already are. A manager who notices things gets faster at acting on what they noticed, which is a real gain. A manager who doesn’t notice now has a fluent, warm, well-structured way to look like they did. The tool didn’t create that gap. It just widened it and made it harder to see from the outside, including from the inside.
The strongest counter, which I think is half right#
The honest objection: a model is a good rehearsal partner, and rehearsal is most of the work for people who find hard conversations hard.
I agree, and it’s how I’d tell someone to use it. Draft the difficult message, ask what’s unfair about it, ask what you’d feel reading it, and then, having found the shape, close the window and go have the conversation in your own words. The value there is that it made you think before you spoke. That’s coaching, roughly, and it doesn’t require the coach to be human.
Where it stops being half right is when the draft becomes the artifact. Then you’ve outsourced not the wording but the noticing, and the noticing was the job. That’s the same trade I got wrong with an AI-written test plan: the output looked thorough enough that I stopped supplying the judgment it couldn’t have.
What’s actually scarce#
The technical floor rose. That’s the whole story of the last three years, and it’s good. What rose with it is the value of everything the floor doesn’t cover.
Knowing there’s a conversation to have. Knowing it’s today and not next sprint. Knowing this person needs a question and that one needs a decision. Knowing when the right move is to say nothing at all and let someone finish a thought. None of that is writing, none of it is in a context window, and no one has automated any of it.
The model will keep writing better sentences than me. It still can’t tell me there’s something to say.
