Two months ago I started writing real code with AI. Today the gap I was hired to fill no longer maps.
That sentence sits oddly. Designers do not usually write production code. Engineers do not usually produce final UI. The split is what most teams are built around. It is what most CVs are still organised by. But in the last two months I have shipped components I would have specced for a developer eight weeks ago, and watched developers ship interfaces I would have drawn in Figma. None of us learned anything new. We just turned on the tools.
This piece is about what the closing of that visible gap actually means. And about a less visible gap that is opening underneath it. Two months is short. I am still wrong about half of this. The half I am sure of is worth writing down.
01The two-month problem
Things are changing in days. Not in quarters. Not in releases. Days.
In April I picked up Cursor. I stayed in Figma for layout, tokens, and component diagrams, but if a component already existed in code I started editing it directly. A week later Codex got good enough to take a multi-file refactor with one prompt and produce a diff that compiled. A week after that Claude Code became my default for sustained agentic work, the kind where I describe an outcome and walk away for ten minutes.
None of these are products I needed time to learn. The interface is a sentence. The job is knowing what sentence to write.
What is uncomfortable is that the gap I was hired to fill, the one between the spec and the working component, has compressed from days to minutes. I have not become better. The tools became better. That distinction matters more than it should.
02What is actually collapsing
Three handoffs that used to define teams have stopped being expensive.
Designer to developer. I now ship UI changes directly. Token swaps, component variants, accessibility fixes. The work that used to live in a Jira ticket lives in a PR. A week ago I rebuilt a small data table from scratch in an afternoon. Three years ago that was a sprint.
Developer to designer. Engineers on my team now produce the first pass of a screen from a sentence. The output is plausible. Sometimes wrong, sometimes the version we ship. Either way it is no longer pre-design.
Researcher to either. Even synthesis is collapsing. A senior engineer can paste twenty interview transcripts into a model and get a clustering that takes a researcher half a day. The clustering is sometimes worse. It is also sometimes good enough.
The handoffs were team contracts. The contracts are now optional. That is the part most teams have not absorbed.
03The Jane Street move
A few weeks ago a Jane Street engineer wrote that they design with Claude more than with Figma.
The headline reads as a stunt. The actual claim is more interesting. They are not saying Claude is a better visual tool. They are saying the medium of design has shifted. The artefact you produce when you describe a screen and let the model render it is not a Figma file. It is a working component in your stack, with your tokens, against your live data.
If that sounds like an edge case it is worth noting that it is happening at Jane Street. The bar for what a senior engineer takes seriously is high. They were not playing.
The work I do is not their work. But the move is the same. The medium is collapsing into the system. The thing you used to draw, you now describe to the system, and the system produces it. The output is closer to production than it has ever been.
04The new gap
Here is the part that took me two months to feel.
The visible gap between specialisms is closing in days. The invisible gap is opening at the same rate.
The invisible gap is between people who know what to build and people who can only execute. Both used to exist. The execution gap was the bigger one, because execution was hard. A designer who could not draw was a designer with a problem. A developer who could not build was a developer with a problem. Execution sorted the field.
Execution is no longer the sort. Direction is. And direction is much harder to fake.
Direction is taste with reasons. It is the call to remove a feature, not add one. It is knowing which of the model’s three plausible variants would get a trader fired. It is a sentence in a brief that closes a hundred wrong implementations.
A model will produce all three variants in seconds. The model does not care which one ships. Someone has to. The someone who can is now the specialism. Everything else is a button on a tool.
This is not a new idea. Every design lead I respect has said something like it for a decade. What is new is the speed. Two months ago direction was something you grew into. Now it is the only thing the role is.
05What this means for hiring
Two patterns are showing up in the way teams are hiring around me.
The first is collapse. A designer who can write code, set the brand system, run a research session, and ship to staging is doing the job that used to take a team. They are not better than they were a year ago. They are amplified. Hiring one of them costs less than hiring three specialists, and the work is more coherent because one head holds it.
The second is widening. Specialists who cannot direct AI are becoming generalists by attrition. Their depth is still real, but the surface a model produces around them is so broad that the depth no longer commands a premium. The opposite is also happening. Generalists who can direct AI are quietly becoming the only specialists left, because direction is the part that does not commoditise.
Neither of these patterns is fair. They are both happening anyway.
06Three things I am watching
Two months is short. These are bets, not conclusions.
Design systems become the contract for AI agents. I wrote about this in April. The argument has only got stronger. A system that holds taste in tokens, governance, and the decision log is now also a system the model can read. A system that lives in screenshots and Slack threads is not. Teams that invest in the first one will look like they are punching above their weight. Teams that have the second will keep wondering why their model output looks generic.
The cost of taste is rising. Taste was always undervalued because execution looked harder. Once execution is a sentence, taste is the part you cannot bluff. The senior people whose taste is calibrated to a domain are about to become extremely expensive. The junior people who have not had time to calibrate yet have the largest gap to cross of any cohort I have ever watched enter the field.
Junior as a tier may not survive. This one I am least sure of. A junior used to learn by doing the execution a senior described. The execution is automated. The learning loop has not been replaced. I do not know yet what replaces it. I doubt it will look like what we had.
07When everyone can ship
When everyone can ship, the question is no longer who can. It is what is worth shipping.
I do not know the answer. Two months is not long enough. What I know is that the people who will figure it out first are the ones who took the tools seriously enough to feel the gap move under them. The rest will be working from a job description that no longer maps to the work.
The gap I was hired to fill is gone. The one underneath it is wider than I expected. I am chewing on it.