Yuval Avidani
Author
"It works, but it's ugly"
Turns out there's one sentence that follows everyone building a website or app with AI these days, and it sounds like this: "it works, but it's ugly." We ask the model to build us a landing page, it hands back something that functions, the buttons are clickable, the form submits, and yet it looks like a generic template from 2015. Gray background, purple-blue button, default font, spacing like someone just threw all the elements on the screen and hoped for the best.
And this is where Impeccable enters the picture. GitHub integrated Impeccable into Copilot, notice: integrated, not acquired, and in my eyes this is one of the smaller-but-smarter moves I've seen in this space lately. Let's break it down slowly, no buzzwords, and understand what this is and why it matters to us.
Wait, what's Copilot and what does "integrated" mean here?
Before we start, two words of background for anyone who doesn't live inside this world every day.
GitHub's Copilot is GitHub's coding assistant. Think of it like "autocomplete on steroids" that sits inside our code editor and suggests full lines of code, functions, and sometimes entire files, as we write. Millions of developers around the world use it every day.
And "integrated," this is the point I want to be precise about, because I've already seen headlines getting confused. GitHub did not buy the company. It connected their technology internally. The difference isn't semantic. In an acquisition, the acquired company gets absorbed. Here, the company behind Impeccable, called Renaissance Geek, founded by Paul Bakaus, remains independent, with its own seed funding from Andreessen Horowitz (if that name means nothing to you, we'll explain in a moment). This is a partnership, not an absorption. This distinction matters because it tells us something about how the industry is evolving: not just giants buying everything up, but small, smart super-layers that plug into the big systems.
So what is Impeccable, exactly?
Here's the clean definition: Impeccable is a "design language": a collection of instructions and commands that makes AI much better at frontend design, for any developer who wants their code to also look good, not just work.
Let's break down the terms, because every word here is doing work.
"Frontend" is the visual side of a website or app, everything we see and click on the screen: the buttons, the colors, the layout, the typography, the spacing. It's the difference between a building's skeleton and the furnished apartment people actually live in.
A "design language" is like a "rulebook" for how things should look and behave. Just like IKEA has a recognizable design language, and Apple has a completely different one, every good system has a consistent set of principles for color, spacing, rounded corners, shadows, hierarchy. A design language is what makes ten different screens feel like the same product.
And what Impeccable actually does is package these principles into 23 commands that the AI can read and execute. Instead of the model guessing what good design looks like, it gets a layer that tells it: this is how you build visual hierarchy, this is how you choose spacing, this is how you make a button look like a button a human actually designed.
Why is "AI that designs beautifully" such a hard problem?
Here it's important to be fair. The reason this hasn't been solved until now is that good design is something incredibly hard to quantify. When we ask an AI to write a function that sorts a list, there's a correct answer, we can test it. But "design me a beautiful landing page"? What's "beautiful"? Everyone has an opinion, and when a model is trained on billions of examples from the internet, it learns the average, and the average of the internet is, well, mediocre.
That's the reason all the vibe coding tools (coding by feel, where we describe in words what we want and the AI builds it) fall into the same trap: they're excellent at getting us to "it works," and they struggle badly to get us to "it's beautiful." That gap, between functional and polished, is exactly where projects get stuck. I've built quite a few things this way, and I always hit the same point: the product runs, but I'm embarrassed to show it.
What Impeccable tries to do is not teach the model beauty from scratch, but give it a "harness." A harness, in AI terms, is an infrastructure layer that wraps around the model, feeds it the right instructions at the right moment, and steers it toward a result. Think of it like a hands-on coach whispering to the model "no, not like that, like this" on every design decision.
Who's behind this, and why a16z is relevant
I mentioned Andreessen Horowitz, or a16z for short. It's one of the venture capital firms (bodies that invest money in early-stage startups in exchange for equity) that shape the tech world. When they put in "seed funding," which is the initial funding round, the "seed money" that lets a young company grow, that's a signal that very smart people are betting on this direction.
And why does this interest me? Because it confirms a thesis I've been carrying around: the next layer of value in AI won't just be bigger models, it'll be thin quality-layers sitting on top of them. Impeccable didn't build a huge new model. It built a smart layer that sits on top of existing models, and by the way, it also works great with Claude Code, not just Copilot. The code itself is open in the repo: github.com/pbakaus/impeccable. I love the humility in that. Not "we invented new AI," but "we taught the existing AI good taste."
Why this is especially powerful right now
Here's the point I'm most excited about. When something gets baked into Copilot, it doesn't reach a thousand developers, it reaches millions, by default, without anyone needing to install anything. That's the difference between a tool you have to go seek out and a quality layer that simply exists the moment you start typing.
Think of it like autocorrect on your phone. Nobody consciously "activates" it, but it changes the quality of every message we send. That's exactly how a default design layer can raise the bar for millions of products being built this year, without any developer suddenly needing to learn color theory.
And there's something democratic about this too. Until now, the difference between a product that looks amateurish and one that looks professional depended on access to a good designer, an expensive and scarce resource. If a layer like this actually works as promised, it shrinks that gap for anyone with an idea and a keyboard.
Bottom line, as I see it
In my eyes, Impeccable-inside-Copilot isn't a "revolution," and it's good that it isn't, I'm wary of that word. It's a smart evolution: taking a real, painful problem ("it works but it's ugly") and giving it a solution that gently layers onto what already exists, without requiring us to change a single habit.
That said, it's right to stay clear-eyed about the limitations. One design language, however smart, carries a particular taste within it, and there's a risk that if millions of products get the same default layer, the internet starts looking too uniform, the way it already happened with certain templates in the past. Good design isn't just "clean and polished," it's also distinctive and full of character. An automatic layer excels at the first, less so at the second. So I'm holding onto my excitement alongside a critical eye.
And still, if this means the baseline of what all of us are building goes from "embarrassing" to "definitely fine," that's a massive gain in a world where most of us aren't designers. It remains to be seen if it actually delivers on the promise out in the field.
So let's think about this together: if the AI we're working with suddenly designs beautifully by default, will that free us up to focus on the idea, or will we stop developing our own eye for good design?