I recently put out a video arguing that frontend development is changing. I don't think the work is going away. The frontend developer title is what's changing, and a lot of us are turning into full stack developers.
The comments had opinions, so I recorded a follow-up.
Pick a stack and get really good at it
One comment that I found really interesting was from a 62 year old developer with almost 30 years in the field. In the comment he described how Claude handles about half his workload now. His advice was to pick a couple of good frontend frameworks like Vue or Angular, pair them with a solid backend like C#, and become an absolute master at them.
I agree with him, and this was something I've always said. It's always good to be at least a specialist in one category. I use Kiro as my agent harness for most of my apps now, and I still want deep knowledge of the stack underneath. That way I can really understand what's happening and help steer the agent when things go wrong. The agents are good. They aren't good enough that I'd stop checking their work.
Is frontend design the safer part of web dev?
This comment is where the title comes from. CRUD apps are the easy wins for LLMs, the argument goes, and attractive, non-generic UI with great UX is where they still struggle.
I sort of agree. With the older Anthropic models you could tell someone one-shotted a design at a glance. It had that very generic Tailwind template look. However, with every new model it's becoming harder and harder to tell. People who aren't looking at this stuff every day mostly think the designs look fine.
I didn't talk much about taste in my last video, but it is important. Looking at a screen and knowing whether the UX works, catching an accessibility problem, noticing that a flow is going to bog users down. That judgment still comes from people with experience. Still, LLMs are getting closer all the time.
I rebuilt Winamp with an AI agent
To test this I picked something with a lot of small UI pieces. I had Kiro Crew recreate the old Winamp MP3 player, and it came back as ErikAmp.
I was really impressed with everything it did with the UI. An equalizer with a row of sliders, a scrolling playlist, play, pause, stop, next and previous, shuffle, repeat, balance, volume. Getting every one of those working by hand would take me a couple of days, maybe less if I did nothing else. The agent basically one-shot it last Saturday. I had it fix a few bugs afterward, and it wrote a whole test suite along the way.
The source code is really simple. I never asked for Vue or React, so it skipped frameworks entirely. It's an index.html, some CSS, a handful of JavaScript files, and a bit of Python.
You could argue that this design was already made, but still. I had it create three brand new themes, and it looks great.
Where it starts to look like AI
The second app I showed is a 5K coach I built to get myself in shape for a race. It tells me what to do each day.
This is about as close as I get to the default Anthropic look. Dark cards, rounded corners, pill badges. It isn't flashy and I wouldn't call it amazing. For my own training plan it's fine, and it took a few minutes with my agent. Building it by hand would have eaten a lot more of my time.
So are we cooked?
I don't think so. Building a frontend is the easiest it has ever been, and I got two working apps out of an agent with very little effort. Deciding whether those apps are any good is still the developer's job, and that takes knowing your stack and having some taste.
I'm also figuring out what to cover next. Do you want more AI workflow content, or more classic Vue and fundamentals tutorials? Tell me in the comments, and tell me where you think frontend is headed.
Top comments (34)
I don't think so as well. Although it is easier to do front-end with AI, it's easy to miss some important aspects of what makes a good front-end. One is accessibility, optimization, etc.
If you prompt the front-end generally, you tend to miss these since it's not really there for you to see. You have to make sure you test the front-end design to see if it's accessible, which can vary depending on who your audience is. You also have to take care of optimization, which is difficult since it depends on browser and one's computer specs.
Even when done with detail prompts, it's a similar experience since you are focus on the looks and not about if it's a good design in general.
Overall, AI makes it easier to create front-end designs, but it's important to catch as many edge cases as you can which is part of the human aspect. Sure, AI can take care of those edge cases, but you never know until you try it yourself.
Great writeup Erik!!
Accessibility is a joke for Opus at least, optimization also not a big deal. FE is solved, that's for sure π₯²
Leave a comment!
Hi, interesting and maybe quite controversial topic.
I want to say, i hope we are not cooked! yet the way we develop is certainly changing. I liked the comments you mentioned above.
AI makes things easier, and can certainly extend one's skills-set a lot. But i love them best for i can learn much faster and quicker.
The line that hit me: "Building a frontend is the easiest it has ever been. Deciding whether those apps are any good is still the developer's job."
I'm a beginner β three weeks into Python, writing tutorials about it. I'm not a frontend dev and I've never built a React app. But this post hit something I've been feeling.
I started learning Python right when AI coding tools got good. And I noticed the same thing: AI can write the code faster than I can read it. That scared me at first. What's the point of learning if AI already knows?
Then I realized the answer is in your post. The agent can one-shot an equalizer UI. It can't decide whether the UX is any good. It can't notice that a flow is going to frustrate users. It can't have taste. That still has to come from a human who's seen enough real projects to know the difference between "works" and "works well."
The other line that stuck with me was the advice from the 62-year-old dev: pick a stack and become an absolute master at it. Not "learn 5 frameworks." Master one. Understand the layer underneath so you can steer the agent when it goes wrong.
That's my plan too. Not to learn everything. To understand one thing deeply enough that AI becomes a tool I direct, not a replacement I depend on.
Great post. To answer your question β I'd love more fundamentals content. The AI workflow stuff is useful, but I need the foundation first.
My thought is very much inline with the intent of your post (with a few exceptions on some specifics) π₯°
There are so many things I have to say on this topic, that I often just don't say much π€£
This is my prediction: AI has the potential to replace all knowledge base workers, with the only exception of really good Software Engineers and Programming Ninjas (I'm separating SWE and Programmers for very good reasons too).
Of course, the exception part is exaggerated to prove a point. But in reality, I too think that AI will NOT replace anyone with deep foundational human knowledge. So, becoming very good at a few things, and having decent amount of knowledge on many other things will be extremely beneficial.
I have a post in my draft that somewhat explains why I think that.
I'll add the link to that post once I publish it.
We need more of these sort of posts to prevent AI CEOs misguiding new generations for their own financial gains. While AI is useful, it can never be replacement for human knowledge and creativity. It can only augment human capabilities.
BTW, before I complete that post, here's a short one: π
Parallels AI coding tools to early calculators
So, We Don't Need to Learn or Use Basic Math / Arithmetic Anymore? π
Really enjoyed this one. The ErikAmp example makes the point pretty clearly: AI is getting surprisingly good at producing the visible part of frontend work.
But frontend isnβt just a collection of screens. Itβs a system of states, interactions, constraints, and user decisions.
Take a checkout page. An AI can generate a polished payment form very quickly, but what happens when the payment request times out after the user clicks Pay? Do we show βPayment failedβ when the bank actually received it? What happens if they refresh and submit again? Does the UI preserve the correct order state? Can they recover without creating a duplicate payment?
The same thing happens with something much simpler like a search box. The happy path is easy. The interesting part starts when the API is slow, returns nothing, returns 500, the user changes the query while the previous request is still running, or the result arrives out of order.
Thatβs where I think frontend engineering becomes more than implementation. AI can produce the screen, but someone still has to understand the state machine behind the screen and decide what should happen when reality stops following the happy path.
AI is making frontend implementation cheaper. It also makes good judgment more visible.
The title change matches what I see. Vue on the front with Laravel behind it is what I write every day, and the work never cared which half I was in that week.
If AI takes anything first, I'd bet on the glue between the two halves: forms wired to endpoints, validation written twice. Deciding what the endpoint should do at all stays with us.
Great read! I think frontend development is changing, but not disappearing. AI can generate components and layouts surprisingly well, yet translating business requirements into a reliable, accessible, and maintainable product still requires human judgment.
In my experience with open-source projects, the hardest part is rarely writing the initial code. It's understanding trade-offs, edge cases, and long-term maintenance. AI helps with implementation, but ownership and decision-making remain human responsibilities.
I've been shipping frontend with AI agents for about 18 months and the "generic Tailwind look" gap is closing faster than people think. Last month I had a designer friend review two landing pages β one I hand-coded, one Claude generated β and he picked the AI one as more polished. Taste still matters but it's becoming a steering skill, not a pixel-pushing skill. The real moat isn't design, it's knowing the stack well enough to debug when the agent confidently breaks something subtle.
The ErikAmp and 5K coach examples make a useful distinction: matching an existing interface and designing a new workflow are different tests. I would separate visual fidelity from interaction fidelity in the player: pause midway, seek, change the playlist, then use only the keyboard and check what keeps its state. For the coach, the equivalent question is whether someone can understand and complete todayβs session without interpreting the whole weekly plan. That gives the agent a concrete user task to improve, rather than another theme to generate.
Some comments may only be visible to logged-in visitors. Sign in to view all comments.