«Vibecoding» is a word that has attracted two extremes. Some hear in it «the neural net writes the code now, programmers aren’t needed». Others — «a toy, it’ll fall apart in production». The truth, as usual, isn’t in the middle but off to the side. Let’s take a concrete task and trace what actually happens and where the time goes.
Take a typical task — ordinary work, the kind there’s plenty of. You need an internal tool: a page where managers see requests, can mark a status, leave a comment, and once a day it exports to a report. This used to be budgeted at a week. Now — a couple of days. Where did the other five go? That’s the interesting question.
Day one: what used to be half the work
The scaffolding that used to eat the first days now assembles in hours. Markup, forms, a sortable table, the database connection, the routine validations — AI writes all of it fast, because it’s been written a thousand times before you. This isn’t creativity, it’s craftsman’s routine, and speeding up exactly that is what produced the «manifold» effect.
And this is where the false impression that it’s done is born. By the evening of day one there’s a thing that looks like it works. It opens, it shows something, the buttons click. An inexperienced eye says «great, ship it». An experienced one knows half is done — and the easy half at that.
Day two: the last 20%
Now comes what AI speeds up far less dramatically, and in places doesn’t speed up at all.
What happens if two managers edit the same request at once? What if the export fires at the moment the database is busy? Who even has access to this page — because an «internal tool» has a habit of leaking outward. What happens to the report if a comment contains a line break or a quote mark? These aren’t flashy tasks, you can’t show them in a demo. But they’re exactly what makes the difference between «works on my screen» and «works in reality, when there are ten users and all of them are live».
AI is a helper here, not the author of the decision. It’ll suggest options, write a snippet for each — but choosing which of them is right for your load and your data, and being answerable for the choice, still falls to a human. That’s what we do together with AI: the boring half goes to it, the responsible half stays with us.
Why it’s still an honest «manifold»
It might seem the trick is exposed and there’s no speedup at all. There is, and a big one — just not where the hype sellers promise it.
The speedup is real across the whole volume of the typical: scaffolding, integrations from documentation, churning through repetitive code, drafts you then edit. On the example task it really is a week compressed into two days. But it compressed thanks to the routine, not thanks to anyone having stopped thinking. The five «vanished» days are the days that went into craft. The remaining two are the days that were always about engineering, and they haven’t gone anywhere.
What this means for you
If someone promises you AI will assemble a complex system whole and without an engineer — that’s a salesperson, not a developer. If someone tells you AI changes nothing and it’s all expensive as before — that’s also untrue, just from the other side.
The truth is boring and workable: part of your task really did get manifold cheaper, part stayed exactly where it was. What sets a good contractor apart today isn’t whether they use AI or not (everyone does by now), but whether they draw the line between those two parts honestly. We try to draw it out loud — tell us your task and we’ll say what compresses and what doesn’t.