Cool. Let's get this going. Welcome everyone. Thanks for coming. My name is Paul and I'm going to try to save you from the midday slump. So, let's see if we can do that.

Okay. So this cold open of this talk will make a lot more sense if you know what Impeccable is. But if you don't know what it is, it's a design skill that turns your coding harness into a better designer and hopefully you into a better designer. And you can try it out. It works across all harnesses: Claude Code, GitHub Copilot, Cursor, Codex, etc. And this talk is really about my approach to how to build this tool as opposed to the tool itself.

Let's get started with a pretty typical website. This is an intentionally very basic website and I want to give you that one as context to what I'm going to show you next. It's an okay website, it's not great, it's pretty bland, but at least it has some color, some something.

Now here is the command "make the workflow section bolder" with Impeccable installed. Now, it's not perfect. There are some problems with this. For example, the section numbers that you have there, that's a very typical AI slop tell that GPT loves to do, and Claude actually loves to do as well, and it sometimes creeps in. So I wouldn't say this is perfect by any means, but I think it's bolder than before.

A lot of people install Impeccable, try it out, and then they ask me: "Well, I don't know, maybe it could be placebo, maybe the model is good enough in itself, right? Not sure if it actually works." Well, let's compare it. This is exactly the same project, GPT-5 on Extra High, same exact prompt without Impeccable. There is a slight difference, I would say, but you be the judge. The before and after is pretty different, and I'm going to talk about why that is and what the approach is.

Designing at the speed of adjectives. To get the elephant out of the room: they are not just adjectives, they're also verbs. But some of them are adjectives.

First off, the role of the engineer and the designer are blurring. I've seen more people become design engineers than ever. Designers are moving into code, engineers are moving into design, and vice versa. Some of the best comments I get from people who walk up to me and say like, "Well, Impeccable gives me this shared language of design that everybody can communicate now in." But really, these worlds are all colliding in my opinion, and the typical handoff is kind of broken. The waterfall process of starting with PM, then going into design, then the handoff and critique process, and then ultimately the engineer builds it—that thing is just not real anymore. I think that's starting to collapse very quickly in small and large companies.

When it comes to the actual work of design, I feel like there's right now kind of two worlds that you can play in.

Either it's direct manipulation sort of like in the pixel space, right? So you're in Figma and you're manipulating margins, paddings, whatever it is. Or you're in Webflow or you're in some other tool where you're directly manipulating the final output or the design. Now that's fine, but then you feel like you're getting behind because everybody else is like agentic and going crazy with loops and loop-maxing, right?

And now you're going on the other extreme, which is simply telling your agent: "Please design this for me" or "Please build this." But it feels like to me at least there's this middle ground that we haven't explored enough. Like what is the exact level of control? How can we insert the human in the loop at the exact right time?

This came to mind—it's a pretty popular tweet a while ago about centering a div with Opus. Yes, not a great use of Opus, but you know it kind of speaks to that problem. It's like: okay, well when do we use which tool? When is which tool adequate to do what job?

In my opinion, right now for most work, the direct manipulation of padding, margins, spacing and so on is actually at an elevation that's too low. And on the other hand, if you go fully agentic, you get web pages like this that you would get in like 2022. Everybody here can point out the typical AI slop patterns. But the purple gradients were actually way beyond that; there are no purple gradients anymore in most frontier models. Now we just get this, and it kind of wears a different suit, as Claude would say. But it really is slop as well, it's just a moving target.

So now we have what I call Claude Beige. We have these Instrument Serif fonts, we have italics. And it's not necessarily a bad design, it's just everything that comes out of Claude design looks that way. And when everything looks the same, you kind of get Algorithmic Uniqlo, as I call it. So also not great.

Now here's my first hot take of today's session: You cannot one-shot design.

Many people want to, and maybe we'll get there eventually. I just don't think it's possible at the current moment, maybe not ever. I think in order to build something that really looks and feels great and is effective design, you need to have something that's context-rich, which has been informed by what you're trying to do, what your audience is, what you're trying to build. And also it will have to be multi-shot. You will have to iterate on your design.

Even if you're a solopreneur—and if you're not a solopreneur, you might have different people who have different opinions, users have opinions—design is messy. It's a process, and wishing it away and just saying like "Solve this problem for me, agent" will not, unfortunately, solve your problem.

Now the other problem is if you don't know the language of design, or if your stakeholders don't know the language of design, it gets really frustrating really quickly too. If you've ever designed something, you kind of know this experience, and the inner feeling builds up...

So really what you have to ask first is: What's the emotional territory? What should this never feel like? What's the reference? What's the audience? These are all really important questions to figure out. You can't go to a human design studio, even if it's the best design studio in the world, go to a design director and say like: "You know what, I want to design for my brand," and the design director nods, says "Got it," and then walks away. That doesn't make any sense. They obviously have to figure out what it is that you're trying to build and what you're trying to get out of it.

Importantly, even on the 2026 version of AI slop—the quickly vibe-coded page—nobody decided anything here. You just one-shotted this, nobody decided anything on this page. It's maybe competent, but completely empty.

So my thesis that I make here, and what I've been trying to build over the last couple of months, is a tool that gives you just the amount of control needed to steer your agent in the right direction with adjectives and verbs.

And these are the types of words I'm talking about: words like make this bolder, make this quieter, distill this (basically simplification), polish this, make it denser, or harden this. "Harden," for instance, would mean make sure that this design works across the board—maybe the performance is bad on certain devices, maybe it's not responsive.

These are words that we use in the world of design. And I noticed when seeing two people attempt the exact same task—one is an engineer who's never really touched design at all, and the other one is a designer—there's a stark difference, even when they use the same model and same harness, in the output that they're getting based on the language that they use. So, I've been trying to put that language, basically compress it into a skill and into a system to be able to express yourselves better.

You could have a baseline, but then you make it bolder and the agent actually knows what "bolder" in your context means. You make it quieter, it knows what it means. You distill, and it knows what it means.

But then: bolder how? What does "bolder" mean in the context of your project? That's a really important question because, as you could tell at the beginning of the demo in the cold open, when you just ask GPT or Claude or whatever to "make this bolder," it doesn't know anything about what you mean. Now it has its own sense of reference of what "bolder" could mean, but it will do whatever it wants, right? It will invent new colors, it will invent new gradients. It doesn't matter what it actually means.

In Impeccable, it's not gradients, not glass, not neon, but actually hierarchy, scale, decisive type—things that don't break your design system and adequately raise the attention that the user would put on it.

This is an actual sentence from the file that is loaded when you say "bolder" in Impeccable: "Show someone your work and say AI made this bolder. If they believe you, you failed." This is something that we actually instruct the agent to do for themselves, and then the agent oftentimes reflects and is like: "I don't know, I think I did some bad work here."

An adjective is a Leitwort. A Leitwort is a German term, and Matt Pocock actually talked about this recently, who also makes some really excellent skills. This is stolen from one of his tweets, but I really think it makes a lot of sense. These are words that you basically imbue with meaning, that you infuse with meaning, and they mean something. They already mean something to the model, but then you translate this into the area that you're interested in, and the right altitude really shifts.

Now, I don't think you can solve any problems at that altitude alone; that's super important too. There is still room for the very beginning—for instance, if you just want exploratory work and you just want to get something on the screen, or you just want the last amount of polish. I don't think right now AI is good enough to replace humans for the last 5%, 10%, maybe even 20% of the work to get it from good to great. But it does actually make it better.

I've built tools for more than 20 years—I guess I'm dating myself here—but whenever I build a tool, I actually map the workflow first. I think about what is the actual workflow of the target user that I'm trying to help. And the workflow of design is not linear, it's messy, has many different stages, there are different people involved.

And so I kind of mapped every part of the process and thought about: What are the injection points along that route and how can we help along that route?

So you start with initialization, with the shaping process of design. You go over to crafting, to iterating with all these verbs and adjectives, and then to the actual harden and polish phase, and finally you kind of bring it back to your design system, you do some tech debt cleanup, design debt cleanup.

Now, here's one example of a command that I wasn't actually sure about shipping, and it's called overdrive. I basically built this command in Impeccable that tries to create something completely over the top, and it was kind of like half of a joke maybe. But people absolutely are rabid about this thing and they love it.

This is not one-shotted. This is a realistic event horizon shader that I've built for my shader library Radiant Shaders. But I did actually use the overdrive command for this to overdrive it even more and make it more ridiculous. So sometimes I come up with these commands I'm not sure about, I test them with the community, and once it sticks, I realize there are more people who are getting joy out of this.

An adjective with nothing behind it is just a nicer prompt. So you really have to tell the agent what you mean by the word that you're using. It has so many different meanings within its training materials that you really have to tell it.

Now here's another thing: I've gotten so many requests for an automatic way of using Impeccable. It's so frustrating. At least once a week somebody tweets at me saying like: "You know what would be cool? Just let Impeccable do all the work by itself. I don't have to decide on any command anymore." That's not the point. The point is to give you a way to steer what you want to end up with. It's never going to be a tool to one-shot design. That's not the intent. There's a pull request standing out right now that I'm going to close because of that. There is no auto, and there will be no auto in my opinion.

This is a bit awkward maybe for some of the next speakers, but I think taste cannot be amplified—sorry, I think taste can be amplified. I call this amplified craft. I think you can work with tools to sharpen your taste and bring out more of your taste. I really do not think it can be lab-grown. Taste by definition is contextual, it's cultural, it's scars. And once everybody replicates the same thing, it starts to become really muddy and we don't think of it as taste anymore. That's my opinion, and I know this is a hotly contested topic. But I'm not trying to solve taste with my tools.

That's what I got. This was about finding the right level of control for design. Thank you very much, and I think I have time for a few questions if you have any. Thank you.