Beyond the Code: What Engineers Really Do When AI Writes It
Will AI replace engineers?
I'd been asked this about thirty times before it hit me: the premise is all wrong.
It assumes an engineer's job equals writing code.
But after all these years, actual coding time only makes up about 20-30% of my day.
What about the rest of the time?
Breaking down PM requests: "They want a notification feature, but who gets notified? When? Will too many notifications get it uninstalled?"
Meeting with PMs to clarify if a requirement is a validated fact or just a hunch from a meeting room.
Drawing flowcharts, looking for what's not written in requirement docs.
Those unmentioned edge cases? That's where problems usually pop up.
This isn't new with AI. Engineers have never just written code.
I recently built my own small product using Vibe Coding, and this feeling is even stronger now.
With AI speeding up coding, engineers have more time for the hardest part:
Deciding what to build, and what not to.
AI Speeds Up the Smallest Part
AI makes development faster. Much faster.
What used to take half a day might now be a working version in 10 minutes. That's a real step forward.
But building Course Kit, I realized something: writing the code is the fastest part; deciding if to write it takes the most time.
Take a notification feature. AI can code it in 10 minutes.
But "does this app really need notifications in the first phase?" I went back and forth on that decision many times.
Adding notifications sounds reasonable, but is it a fact or a feeling?
Is there data showing users miss classes without notifications? Or do I just feel they will?
Will notifications annoy users? Without them, will users forget classes? Is there a lighter approach than full notifications?
This is what thinking really looks like. Not zoning out, but constantly running "if...then...but what if...?" loops in your head.
Adding a feature is easy; deciding if it's worth adding is what takes time.
AI can suggest solutions based on your product background and development status. Pick a solution, and it turns the idea into code.
But whether that idea is worth becoming code? You still have to decide.
Ultimately, an engineer's core skill is distinguishing facts from feelings.
Is this a user-reported problem, or something everyone feels should exist in a meeting room?
AI can't do this. It can't tell what's validated and what's just a feeling.
Same Caution, Different Nuances
And facts look completely different across industries.
I've worked in Fintech and on IoT projects for medical measurement.
Both industries require careful data handling. But careful means very different things.
In Fintech, you fear miscalculated money. Every transaction record must be kept, traceable, auditable.
Lose one record, and you might have an unaccounted transaction.
In medical IoT, you fear data affecting diagnosis. A three-second measurement delay might not just be a bit slow.
It could mean the entire dataset is unusable. Where patient data is stored, how it's transmitted, who can see it — each step has different regulations.
Even with the same data processing, the resulting product can be completely opposite.
This ability to know what to be careful about in a given context isn't learned from documents.
It develops slowly, from being in that environment, making mistakes, and thinking things through.
AI can write code for you, and suggest options applicable to your industry and situation.
But your judgment comes from your journey, not from which language you code in.
After AI Codes: The Engineer's Real Work Starts
When building products with Vibe Coding, a common scenario is: AI generates something that looks like it runs.
Open it up, and the functions work, the UI looks right. For a moment, you think, "Wow, it's done!"
But your instinct tells you: It runs, but does it run correctly?
Will this logic hold up under heavy data loads?
Does this implementation leave security vulnerabilities?
What if a user operates it in an unexpected way? Will it crash?
This is the same looking for what's not written in requirement docs mindset, now applied to AI-generated code.
You find overlooked scenarios, then throw them back to AI to discuss what else I haven't considered. (It's quite good at this, since imagination is its strong suit.) (Laughs)
AI is like a fast-typing teammate. You don't accept its output wholesale; you review and commit only when it's solid.
This isn't distrusting AI; it's just what engineers do: code review.
AI writes the code, but you still decide if it goes live.
So, will AI replace engineers?
I think a better question Is your accumulated judgment, finally unshackled from typing speed ?
Writing code has always been a small part of an engineer's job.
AI speeds up that small part, but all the other things that truly require breaking down, judging, and deciding — none of those have disappeared.
If you already feel that writing code is just a small part of your job, then congratulations, you're already prepared.
If you're also thinking about building your own stuff with Vibe Coding, let's chat.
