A year ago a new subject area took us weeks. Now it takes about a day. People assume we found a better model. We did not. What changed sits one layer down, and it is much less exciting than AI, which is probably why nobody talks about it.
What a vertical actually is
For us a vertical is a complete subject world for a specific age group. Space for ten to thirteen year olds. Science with virtual labs. Each one has its own lessons, its own mentor, its own illustrations, its own tone. It is not a skin on top of the same product. That is exactly why it used to be slow.
The design system is the reason, not the model.
The bottleneck was never the writing
You would expect the content to be the hard part. It is not. The hard part was that every new vertical meant new screens, new components, new spacing decisions, new edge cases, and a designer and a developer arguing about a card layout for two days. Ten decisions per screen, forty screens, and none of it visible to a child.
So we killed the decisions instead
We built the design system properly. Real components, real tokens, real illustration rules, real page patterns. Not a style guide in a document, which nobody reads. Components in code that are the only way to build a page. Once that existed, building a vertical stopped being a design project and became an assembly job. And an assembly job is something AI is genuinely good at.
This is why AI-assisted building fails for most teams
Point an AI at a codebase with no system and you get a beautiful prototype and three weeks of cleanup. The output looks right and is subtly wrong everywhere, and fixing it costs more than building it by hand would have. Point the same AI at a codebase where there is exactly one correct way to build a card, and it builds the card correctly. The constraint is what makes it work. People keep trying to remove the constraint to go faster, which is precisely backwards.
The part nobody warns you about
The design job changes, and not everyone enjoys it at first. Our designers stopped drawing individual screens. They now own the system, review what comes out of it, and decide what is allowed to enter it. That is a more senior job and a more interesting one, but it is a different job from the one they applied for. If you do this, say so out loud to your team before you do it, not after.
What it does not fix
Speed does not make a thing worth building. We can ship a vertical in a day and we still argue for a week about whether a subject deserves one at all. The bottleneck moved from building to deciding. That is a better bottleneck to have, but it is still a bottleneck.
Take these with you
- 1The design system is the reason, not the model.
- 2AI-assisted building only works where there is exactly one correct way to build something.
- 3Removing constraints to go faster makes you slower.
- 4Designers move from drawing screens to running the system. Warn them first.
- 5When building gets cheap, deciding becomes the bottleneck.
