)
)
TL;DR
One year ago, our product design team vibe coded a whole new 3 Sided Cube website in a week. Since then the tools have gone from "build a web module" to "build almost anything." The biggest unlock wasn't a flashier model. It was a plain text file called design.md. Then, the most useful lesson was that AI got fast at making things and stayed slow at knowing which ones were any good, and that call still lands on a person.
Tune in for the rest of the Factory Rebuild:
Key takeaways
Vibe coding grew up. In a year it went from "build a web module" to "build almost any app, tool or plugin."
Keep humans in the loop where correctness matters. Security, accuracy and genuinely new ideas are still a human job.
Fewer, smarter prompts win. Multi-agent loops beat brute-forcing 162 prompts at a single screen.
AI has no taste or opinion by default. Feed it real context and human judgement, or it produces work that looks like everyone else's.
Design.md files are the unlock for designers. A plain Markdown design system your AI agent can read is the cleanest way to keep AI output on-brand.
More output is not more value. The bottleneck has moved from making things to reviewing and focusing.
The experimenting never stops. From agents and loops to plugging Claude into 3D software, today's tinkering is next year's workflow.
Back in 2025…
Roughly a year ago we did something slightly mad. We popped our product design team in a room and tried to rebuild the entire 3 Sided Cube website in a week, using AI to do the heavy lifting.
We were vibe coding before we even knew that's what it was called (the ultimate flex 💪)
At the time, it felt like pure magic. We were building live modules, watching each other's cursors move around a shared canvas, handing polished front-end off to our developers. Our minds were truly blown.
“The days of using V0 to build little modules and hand that off to the developers feels quite quaint.”James Marriott, Head of Product Design 3 Sided Cube
Fast forward twelve months and a LOT has changed. First off, the 2025 version of us looks adorable. But the interesting bit is what got better, what got worse (yes, worse), and how we'd tackle that same brief today.
So we sat down with James (our Head of Product Design) and Jake (one of our Senior Product Designers) to talk through it all!
)
What is vibe coding for designers?
This term gets thrown around a lot, so let’s set the record straight.
Vibe coding is describing what you want in everyday language and letting AI build it. Screens, flows, states, sometimes the code itself. For a designer, it shrinks the gap between "I've got an idea" and "I'm prodding at a working version of it" from weeks to minutes. No handoff and no waiting on a dev to pick it up.
That's the idea, anyway. In reality, the process has a lot more human fingerprints on it than the word "vibe" lets on, which is roughly what the rest of this blog is about.
We can build (almost) anything now
Last year we were living inside a vibe coding walled garden. Our tool of choice back then could build websites and web apps, and that was it. Clever, but we spent a lot of energy trying to work around the walls.
Now it feels like we're frolicking in open fields with a couple fences here and there.
"It feels like in a year it's come out of the limitation of being a web app. We're now able to literally create anything."
Jake Bishop, Senior Product Designer, 3 Sided Cube
The team has built Mac apps to speed up fiddly desktop tasks, Figma plugins to smooth out our own workflows, and apps that run properly instead of the sad half-working kind from twelve months ago.
Another point that shows the changes perfectly was from a recent creative review sesh. One of our other Product Designers, David, was showing off some mobile designs when, halfway through, the rest of the team clocked that it wasn’t a flat mockup. Instead, it was running on an iOS simulator and behaving like a real app. All because David wanted to see the designs live after designing them in Figma. Of course, it would need a look from the developer side of things, but it just goes to show the possibilities.
A year ago, that would’ve been a thing of science fiction!
Steering AI, not driving
Last year, one of our best website modules took many, many prompts. Back and forth, back and forth, dragging it towards what good looked like.
“Before, to get something that stood out from the rest of the AI stuff, you really needed to constantly prompt. I'm pretty sure there were 162 prompts on one of them. Now I'm building full apps and you don't need to do as many.”
Jake Bishop, Senior Product Designer, 3 Sided Cube
Now, Jake builds entire apps with far fewer prompts, which is much better for his sanity. The difference isn't that he's stopped prompting, it's actually the type of prompting.
Part of that shift is multi-agent workflows. This is where, instead of one prompt trying to nail everything, you send several AI agents off to work on different parts in parallel. LLMs tend to do a 70% to 80% job on the first pass, so this loop closes the rest.
"An agent goes and checks the discrepancies, fixes them, checks again, and keeps going until it genuinely feels aligned."
James Marriott, Head of Product Design 3 Sided Cube
It's a completely different rhythm to the old linear cascade of "designer prototypes, developer investigates, handoff, wait, review, repeat".
)
Where AI still trips up
Every vibe coded website looks the same
AI doesn’t have a creative opinion because it runs on data. So when you ask for the ‘best’ version of a design or feature, what you’ll get is actually the most common one. If something’s used frequently, it has to be good, right?!
So, left to its own devices, AI makes everything look identical.
We learned this the hard way on last year’s website build. The first design drafts were fast and impressive, then we noticed it kept producing the same structures.
"It will take whatever shortcut it can. If you don't give it specific context, it takes the path of least resistance, and it will look like every other thing that's out there."
James Marriott, Head of Product Design 3 Sided Cube
Everybody's vibe coded website looked like everybody else's vibe coded website. Take a look at recent AI related startups and you will start to see the same gradients and three-card feature grids popping up everywhere.
“It really didn't have a creative opinion. It was just, this is what people do for a card system, we'll shove that in the website. And everybody's vibe coded websites were the same.”
Jake Bishop, Senior Product Designer, 3 Sided Cube
Once you've seen it, you can't unsee it. And if your product looks like every other AI-built product, you've lost the one thing that made someone stop and remember it.Don’t worry! There’s a fix for this, the Cold Brew method. But more on that later in the blog.
Claude thinks everything it builds is fantastic
Taste and intuition are high up on this list. When James works in Claude, he can watch it review its own interface. It takes a screenshot and, without fail, announces "this looks fantastic." Every single time. Then you look at it and, well, it doesn't. It thinks it's nailed it and it rarely does. Gotta love Claude’s self-confidence though.
New ideas are still ours too. AI is trained on what came before, so it's brilliant at the known pattern and useless at the genuinely new one. We've been updating one of our apps for Apple's new Liquid Glass look, and the AI can read the guidelines fine. What it can't do is invent clever workarounds for a situation nobody has documented yet. That takes a human going "try it this way round." The good solutions came from us handing it knowledge it doesn’t have.
Hallucinating quotes
Picture this, a detailed planning meeting, a proper run-through of how we'd architect a whole new process in Jira. We got our AI note-taker to summarise it, but what we didn’t ask for was for it to put words in one of our Cube’s mouth. It completely hallucinated Kev saying "I think we're all farts."
Kev did not say that. Not even close. Yet, it was delivered with AI's usual complete confidence.
Funny, in this case. Less funny when the stakes are higher. So let this be your reminder: anything that has to be correct, secure, or deterministic needs a human firmly in the loop.
Using AI to undo AI
Initially, AI helped tidy rambly thoughts into clean Slack updates. But now half of our messages have been inflated into cluster bombs of bullet points, most of them not relevant to the person tagged.
"Every message is turned into a cluster bomb of bullet points, half of which aren't relevant to me even though I'm tagged in it. I'm having to use AI to summarise what someone's used AI to make longer"Jake Bishop, Senior Product Designer, 3 Sided Cube
And we’re not alone. This is the problem almost every organisation is wrestling with. AI gives equal weight to everything, whereas a human knows what actually matters in the moment for a specific message.
So our fix is to put the human friction back in. In the places that need care, we stand the AI down a little and let people do the thinking. Think of it as a bit of course correction, rather than complete retreat.
)
The tricks that made it work
The real breakthrough was a text file
The biggest leap for us as designers wasn't a smarter model or a shinier canvas. It was design.md.
If you haven't used a design.md file yourself yet, it's a plain Markdown document that captures all of your design system in one place.
Spacing
Typography
Colours with exact values
Components and their variants
The reasoning behind design choices
Then your AI coding agent reads it and applies it consistently across whatever you’re working on. Markdown happens to be the format Large Language Models (LLMs) read best, so there's almost nothing to configure. You just point the agent at it.
Last year, keeping AI on-brand was a slog. We'd build a design system, ask the tools to follow it, and they'd still wander off in a million different directions. Now, even a low-effort design.md gets our 3 Sided Cube branding replicated. And the more time you spend on the file, the better the design output gets.
"The markdown format is what LLMs really thrive off in terms of interpreting context. So it just makes it that much easier for them to pull that into whatever it's designing."
James Marriott, Head of Product Design 3 Sided Cube
If we'd had this last year, we'd have skipped a mountain of headaches and a lot of "design in Figma, then rebuild it by hand" busywork.
Jake’s also found another way to stop AI from going rogue. A private component library published as a Node Package Manager (NPM) package works because AI can't edit the library, it can only reference it. So, every button in the app Jake built pulls from that one locked source. No drift. No "let's just invent a new button and change it everywhere." A small, sharp trick that saved him literal hours.
The Cold Brew Method
How do you stop everything looking the same? You make sure AI gives you options and you stick a human in the middle to choose between them.
Our workaround got a daft, but very well fitting name: The Cold Brew Method. It actually came about after Jake had hit a ceiling with the website designs. He wasn’t getting any further and his frustration was building. So, fuelled by cold brew coffee, he went back to the AI drawing board.
The idea was to wrangle the AI into producing loads of genuinely different variants in one sweep, then pick the strongest. We built little tools with sliders and exposed properties so we could easily tweak a design ourselves instead of re-prompting into the AI void.
Now it’s funny to see that those homemade hacks are now just how the AI design platforms work. For instance, Claude Design now offers up variants pretty much straight away and provides panels so you can tweak certain elements using sliders and dials, rather than only prompting in conversation.
"Claude Design has the feature where it exposes those little panels so you can tweak the dials and see the variants, and it will, in conversation, ask you about the different variants you might want to see. Very similar to Cold Brew."James Marriott, Head of Product Design 3 Sided Cube
The scenario of hitting a wall, thinking "if only it could do this", then a few weeks later something ships that does exactly that, is happening really regularly now. It just goes to show how fast the world of AI is moving.
)
What AI's doing to us
Everyone's becoming a "product creator"
The biggest shift of all is the role itself. At Cube we're moving toward a more merged way of working, where designers touch a bit of development, a bit of product ownership, a bit of everything. The industry's calling it ‘the builder’, or ‘the product creator’. One person, super-charged by AI, being involved at pretty much every stage of the build.
Which raises this question: if anyone can build now, what's left for designers?
From our experience, the answer is loads! Yes, the building got easier, but deciding what to build, and whether it's any good, didn't. AI still hasn’t got a clue about user experience. That part's still a designer’s job, and it matters more than ever because it is the exact thing that keeps users coming back.
There's also a lovely side effect to this merged way of working. A deeper appreciation of the whole development process.
"It builds a lot of empathy. You learn a lot by osmosis. When you ask for something in a very vibey way, you see the process Claude is breaking it down into." James Marriott, Head of Product Design 3 Sided Cube
By vibe coding real fixes into real apps, our designers have picked up what a pull request (PR) is, why merging really matters, exactly why a vague ticket bounces back. Jake spent a year learning and now he writes tickets that developers accept first time, because he properly understands what information developers need.
“I'd heard 'PR' my whole time working here and I was like, yeah, it's some magical thing. Merging was a whole other thing I decided for ages not to learn, because it sounded too complicated. Now I understand how it works."Jake Bishop, Senior Product Designer, 3 Sided Cube
Focus is the new scarce resource
When you can build anything, you try to build everything, and it can get overwhelming very quickly. James has explained how he felt when faced with all the new AI possibilities.
Jake also has 39 side projects on the go, which sounds cool but is actually pretty draining. All those projects pull you in so many different directions. We’re finding that focus is the new scarce resource. Not capability. Being able to work on so many projects and possibilities is amazing, but the important thing is not working on all of them at once.
"It's like, my God, I can build anything. And then it's my God, I can build anything, and it becomes quite overwhelming, where to pick a focus."James Marriott, Head of Product Design 3 Sided Cube
The ‘build everything’ compulsion is something our Head of People, Emillie, is very conscious of. She’s our very own Gandalf, protecting Cubes’ time, setting realistic expectations for using AI, and moulding our culture as we go.
)
How would we rebuild the website today?
It depends. We know, the most annoying answer ever, but it’s also kind of the whole point.
Some things we know for sure, like:
You wouldn't catch us using last year's tool.
We'd send agents off to research and build different parts in parallel.
We'd lean on a strong design.md from day one instead of designing in Figma and rebuilding by hand.
There’s no way we’d skip the human part like spending time mood-boarding, pulling look-and-feel apart, and iterating on judgement, all before a single agent got involved.
The projects that turn out well are the ones where a person made the hard calls first and let AI execute. Skip that, and you get the same generic output as everyone else.
So what’s next?
If a year took us from "build a web module" to "build almost anything," we're not about to stop poking at the edges!
Jake's got his eye on going deeper with agents and loops, getting AI to juggle more at once. His current rabbit hole is connecting Claude to Blender, the 3D software, to see what happens. So watch this space.
James is focused on growing into the product-creator role properly, stretching his technical range and getting comfortable wearing a bit of the developer hat. Part of what's making that possible is the sheer number of MCPs (Model Context Protocols) being released. Every new one leads to more possibilities.
We haven’t figured it all out yet. The tools will keep changing, faster than ever.
The bit that stays the same is us. Curious, slightly over-caffeinated, and still the ones deciding what's worth building.
Straight from the Cubes
Want to hear more of the low-down on how the factory rebuild has gone across the teams at Cube? Catch up with our fellow Cubes here:
Disclaimer: We used AI in the process of writing this blog, but a human was still at the centre! We'd be poor advocates for going AI native if we didn't use AI to help write about it. The story is ours, the quotes are our team's own words, and every line got the human pass 💚
FAQs
What is vibe coding for designers?
Vibe coding is describing what you want in plain language and letting AI build it, from screens and flows through to working code. For designers it collapses the gap between an idea and a testable version of it, often to minutes.
What is a design.md file?
A design.md file is a plain Markdown document that holds your design system, including colours, typography, spacing and components, in a format AI coding agents read easily. It keeps AI-generated interfaces consistent and on-brand across a whole project without endless re-prompting.
Does vibe coding replace designers?
No. In our experience it removes grunt work and speeds up prototyping, but taste, intuition, original ideas and the hard product decisions stay with people. AI drafts. Humans decide.
Is vibe coding actually good for real products?
For prototyping and internal tools, it's excellent. For production apps it works well with a human in the loop and developer review, and it should be kept away from anything that has to be perfectly secure or deterministic.
What's the "product creator" or "builder" role?
It's an emerging role where one person, amplified by AI, spans design, development and product ownership rather than sitting in a single lane. We're moving toward it at Cube, one learned pull request at a time.
Published on 28 August 2026, last updated on 28 August 2026