Writing was never my job. I build things. But AI nudged me to put words down, and that small shift says a lot about what's changing for engineers, and what isn't.
Writing wasn't my job
Some context. I've been Head of Frontend Engineering at GoodBarber since the company started in 2011. I build the rendering engines that turn no-code projects into native apps. I teach on the side, at the University of Corsica, and I occasionally get on a conference stage. I knew how to build and how to explain out loud. Writing is another matter. I did publish a few technical articles over the years, but each one cost me an unreasonable amount of energy, and none of them made me want to write the next one. Regular, public writing, with my name on it, was for other people.
Not that I had nothing to say. After all that time inside the same product, you have opinions on everything. The story wasn't the problem either: structuring, sequencing, arguing, that's my daily work. What cost me was building the sentences. Finding the right phrasing, redoing it, redoing it again, feeling it was heavy without knowing why. At that price, I always had something better to build. Almost every time, the page lost. For fifteen years.
What changed is very practical. The AI builds the sentences, and I do everything else: the ideas, the order, what's true, what's false, what stays, what goes. It transcribes, I review. And reviewing is what I do every day with code. The fixed cost collapsed, and it turned out the ideas had been waiting behind it.
And ideas weren't the only thing waiting. There are also affects: the attachment to a product you've been growing since 2011, the pleasure of an architecture that finally holds. I had never managed to get those across, because I lacked the technique. I don't lack it anymore. A bit as if I had learned to paint in a few months.
What AI changed in how I build
I trusted that shift because I had already lived it on the engineering side.
The obvious things accelerated first: boilerplate, scaffolding, exploration. Trying three architectural approaches used to cost three days, so we argued in meetings about which one to build. Now you build all three before the meeting and argue about results. Everyone knows that part of the story by now.
The deeper change is where my hours go. My time has moved from writing code to deciding and verifying. The machine produces, and I judge, check, accept or reject. A concrete example. I led a feature that lets an AI agent design an app's backend: a real Supabase database, live, with the user's actual data. The interesting engineering wasn't the generation. It was the guardrails. Destructive actions were gated on reversibility. Every proposed operation was logged before running. External calls were fenced to an allowlist of domains. And one principle that sums it all up: real data or an honest empty state, never an invented number.
That sentence isn't code. It's an editorial judgment about truthfulness, applied to software. The machine took the typing, and what was left of my job was deciding what should be allowed to happen, and what "honest" means in an interface. Somewhere in there I noticed that the two halves of my week, building with AI and writing with AI, had become the same activity. The machine handles the transcription. I supply the judgment.
Why an engineer still has things to say
Which raises the fair question. If AI writes fluently, why would an engineer bother writing at all? Won't the machine say it better?
It will say it smoother. It can't say it at all, in the sense that matters, because the value of an engineer's writing was never the prose. It's the scar tissue. The machine can write a beautiful paragraph about any technical subject; it wasn't there for any of the decisions that aged badly on us, or for what it took to walk them back. A position held for fifteen years isn't retrievable from a corpus. It has to be reported, by someone who was in the room. And what that person reports isn't just facts: it's what they felt when it broke, what they believed before they understood, the point of view of the one who carried the decision. Ours, for example: at every generation of hybrid frameworks since 2011, we made the same call, native. A call held against the fashion of the moment, argued for years. Many competitors gave up along the way; we're still here, running on engines the years have made solid. The machine can summarize the ten thousand blog posts of that debate. It has no point of view.
That's also why handing the sentences over doesn't bother me. For a novel or a short story, I would understand: there, the sentences are the work. Here I'm not making art, I'm sharing a point of view. And it's judged on what it claims, not on the sentences that carry it. Sentences matter the way technique matters in painting: without it, nothing gets through; but it isn't what people come to see.
The best objection I've read comes from Ted Chiang: art results from a quantity of choices, he writes, and a ten-thousand-word story represents about ten thousand of them. Someone who finds formulating sentences a chore isn't really working in their medium. He's right, for art. But my medium isn't the sentence, it's the point of view. An engineer's essay is judged like an app, on its pertinence, and my ten thousand choices were made elsewhere: in fifteen years of production. On the essential point, though, we agree: what makes a text new, he concludes, is that it comes from a unique life experience.
My colleague Muriel crossed the same bridge in the opposite direction. She works in marketing, and she taught herself to build with AI. She got the speed, and then she hit the part I take for granted. In her words: "the machine hands you something that works, and you couldn't explain why." Read together, her story and mine say the same thing from two sides. AI moves the execution, and the judgment stays with whoever has the experience.
The balance, as I see it
Here is what this small detour taught me about the question everyone is asking: is the machine replacing us?
AI didn't replace the part of my job that mattered. It exposed it. Take away the sentences, code or prose, and what remains is judgment, experience, and the will to say what you actually saw. The machine amplifies a voice. It doesn't contain one.
I'm still at the same company, teaching students who will never know the blank-page tax. The machine takes on more of the execution every month. And what's left over looks, each time, more like the reason I got into this work in the first place: deciding what's worth building. And now, sometimes, what's worth saying.
Mathieu Poli — Head of Frontend Engineering @ GoodBarber. I teach and write about frontend engineering, product design, and AI — and everything that happens when the three meet. · X: **@hellomathieup**
Top comments (0)