Hello everyone. Quick check, how many Brits do we have in the room? And a little whoop whoop.

None? Great. Okay. >> [laughter] [gasps] >> So, I'm Harvey Whiteing from Meta Design and I'm here to talk about something that everyone in this room is quietly wrestling with, AI slop.

A few months ago, that's exactly where we were.

We had all these incredible new models. They had the knowledge of the entire internet, but they didn't have taste.

They didn't know how design systems work. They didn't know the difference between close enough and right. And the challenge at Meta is that close enough doesn't cut it.

We serve millions of businesses and billions of people. Every surface, whether it's Instagram or Ads Manager, has to rise to the bar of our design system.

And that's just not something AI can do out of the box. But the question is, why not?

And today, I'll take you through that journey and what it might mean for your team's path forward.

We'll start at the beginning, a partnership between Figma and Meta Design to improve Figma makes ability to work within a design system.

And then we'll look at the progress, warts and all, from wobbly early iterations to results that approach our quality bar.

And finally, how that can lay the foundations of this agentic era of design. But first of all, some context on why this matters right now.

We're in the middle of an explosion. Vibe coding has gone mainstream. Peo- people are building entire apps from just a prompt.

The AI design tools are shipping faster than anyone predicted. And every week there seems to be a new tool that promises to turn a napkin sketch into production UI.

And the energy is incredible. But here's the thing no one is talking about.

Most of what these tools produce doesn't hold up at scale. It looks plausible. The layouts make sense. The design spacing feel close. But when you put it next to your actual design system, the gaps are [snorts] everywhere.

There's missing states, components that don't even exist, content that breaks every guideline that you have.

This explosion of AI coding is the most exciting, but also the most dangerous era for AI and design tools.

The question isn't whether AI is part of the workflow, it already is. The question is whether we let it ship slop, or we teach it taste.

And that is what we set out to do. So let's talk about the problem concretely.

This rapid acceleration of AI design tools has created a new category. Fast, but wrong. And we started experimented with AI generated design for odds products.

And regardless of the model or the tool, they all started failing in the same ways.

Low fidelity, wrong typography, inconsistent color, missing or just malformed components, things that would get flagged in any design review.

The second is around low repeatability. The same prompt gives you a different result every single time. And that unpredictability erodes trust.

And the third is context collapse and complexity. A single form, AI does okay. A full odds manager experience, and the quality falls off a cliff.

And that last piece creates this new adoption barrier. Designers will not use a tool where speed comes at the cost of quality.

Just to reiterate, designers will not use a tool where speed comes at the cost of quality. They just won't.

And here's the paradox that kept us up at night. AI can generate design faster. But without quality control, that just means more rework for your designers.

More confusion for your engineers. And more cost for the company. Speed without quality is just expensive noise.

So, to unlock AI design, we have to solve quality first. So, why was the agent dropping the ball? We dug in and we found three root causes.

First, proprietary dependencies. Our products use internal code libraries, and agent trained on the open web can't bridge that gap. It doesn't know what it doesn't know.

The second was our design system's complexity. We're not just talking about a handful of components.

We're talking about hundreds. We have sprawling primitives across type, color, spacing patterns, interactions, content design. There's do's and there's don'ts.

Sometimes there's intentional friction between simplicity and sophistication. And that's hard for a human to hold in their head.

Let Let alone a model. And the third one was just unrealistic expectations. We were expecting AI to one-shot something that even an experienced designer would iterate on several times with multiple rounds of feedback.

If iteration is how designers get to quality, what if we gave AI that same opportunity? What if we got scientific?

So, we started to set up things like golden sets. Golden sets, if you haven't heard it, it's a fancy word for lots of test cases.

And if you run your system multiple times over and over through those same test cases and get really scientific with like how do you measure that?

How do you put a metric to it? And I've got some examples here. Then you can get feedback. You can re-prompt.

You can isolate exactly where it went wrong and you can get it back on track. And we also set a goal post. If we can produce prototypes that at least 80% high accuracy, and by high accuracy I mean perfect.

And if it utilized every single component then it then it that it should, then we'd say it generated a sufficient quality to take this seriously.

So, with this technique, we were able to rebuild our components in a pixel exact way. Not close enough. Exact. And the model could learn from its mistakes within a session and course correct.

We took those iterated upon components and we added strict guidelines for the agent to follow and how it should use them.

All captured inside a Figma make template. And it started to produce the results we'd wanted all along. That got us to the bar we'd set for our pilot.

87 80 7% in fact above where we'd set our bar. And when I say high accuracy again, I mean it's free of issues.

It's production ready. And that's a massive jump from where we started.

You can see here on the left. That's our target. That's the design we started with. And on the right you can see our iterations.

And it starts off pretty wobbly. But in the end the left is the right. We got there. We were able to produce exactly the quality that we wanted.

But, these manual feedback loops, they don't scale. So, what we really needed was a context delivery system that could operate at meta scale.

So, we started developing what was then a new protocol, MCP, model context protocol.

And we investigated a fundamental investigated fundamental question, could we capture all dimensions of our design system in this infrastructure?

The color, the type, the spacing, the components, not just the components themselves, but the documentation around them. Our content standards, our illustrations, all of it.

So so that any AI tool could get the exactly the context it needed, exactly the moment it needed it.

And we partnered with Figma, and we integrated this Figma make for the first time. And at the time, this was a net new capability that we codeveloped.

And the results were immediate. Design system aligned output straight off the bat, in one shot, without even needing a template.

And the model could pull in exactly the context it needed, at exactly the moment it needed it. And we could actually push beyond the component code into areas like content standards, where we just know we need a vast amount of context and judgment to do a good job.

And it's so exciting to see connectors and MCP connections and are broadly available to everyone in Figma. And hopefully you understand why this is so critical to pushing forward quality in what we do with AI.

So, what's next? What are we working on now?

What's what's the new latest and greatest? We see a layered stack, a path forward for the entire industry.

The foundation is context. We've talked a lot about it today, and for good reason. If you get nothing else from this get this.

You must get your context foundations right. Everything that comes next depends upon it. If your design system isn't structured in a way I way that AI consume, you are building on sand.

And the next layer is skills. If this works.

Spoiler alert. Um and the next layer is skills. And so, this is also something that we're launching in in Figma natively.

And context is knowledge, but skills are knowing how to apply that knowledge. And a model that knows your color tokens, it's useful.

But a a model that knows exactly when to use them and why is powerful. Next slide, please.

Awesome. And this final layer, as you guys all know, is agents. So, agents bring this all together. You take the skills, you take all of that foundational context, and when you put it together, some magical things start to happen.

So, this is what gets me excited. And as we all know from yesterday, it takes a lot to get a British person excited.

It's genuinely about expanding what's possible. You can give an agent a task in the evening and get results in the morning. You can jam with the agent, expectation align, form a plan.

And let's say you decide you want to do 10 different directions. Well, as you can see, you can have an agent working on every single screen, and we're able to just with six words or so produce a design for an entire feature.

You can just have the seed of an idea and then let it grow into a full concept and then iterate on it. Give feedback, push it forward at a pace that was never possible before.

The designers on our teams aren't doing less, they're thinking bigger.

So, what does this mean for you? I I wanted to give you guys a bit of a cheat sheet and how you can be successful in this new era and just give you three things you can take home today.

Number one, audit your context. Look at your design system and ask, could an AI consume this?

Are your tokens documented? Are your components structured with clear rules? Remember that anything that is not specified is randomized or aligns to some web convention that you didn't define or agree upon.

This is the foundation of that everything else will build upon.

This is your number one priority. Number two, embrace the feedback loop. Don't expect one-shot perfection from AI.

Treat it like a collaborator. If a new designer started on your team, you wouldn't just say, "Hey, go do this." You'd say, "Hey, this is my customer.

This is the problem we're solving. This is all the context you need and this is how you move forward." So, you can re-prompt.

You can correct. You can iterate. That's not a workaround.

That's the workflow. And these models are trained on imperfect humans, so they're imperfect, too.

And lastly, prepare for agents. These tools are arriving faster than any of us expected. The teams that have their context and their skills ready to go, you will be the ones that move the fastest when agents become production ready.

And so, we went from AI slop to quality at scale, not overnight, through deliberate, unglamorous work on our foundations.

And every team in this room can start on that journey today.

So, I'll leave you with this. We started talking about AI slop with outputs that were fast, but wrong. That eroded trust.

That made designers skeptical and honestly, that skepticism, it was earned. But look at where we are now. Design intent flowing from design to prototype to the product itself and back again.

Not perfectly, not everywhere, not yet. But the foundations are in place. The context, the skills, the agents, they're real and they're getting better every week.

We came to realize when all of this comes together, it actually creates a paradigm shift.

Where we're headed, design is becoming truly multimodal. It turns out these language models are amazing at translation.

If it's Swift, HTML, Figma design, they're all just code. Code that can be translated from one medium to another with with little to no context loss.

You can design in one medium and translate to another. You've got a web prototype and you want to bring that back into Figma design, you got it.

You want to take that design, gather a bunch of comments and send these changes back into your prototype, easy.

The tools are starting to catch up with our ambitions. If all the lines in this visual, instead of being wobbly, they were straight, then you could just pluck a version of your production product like a ball of clay, you could give it shape and you could just throw it back into the product.

And I don't think it's impossible. In fact, I'm seeing it working at Meta already and we're improving it every day.

And so, the teams in this room, the ones building design systems, the ones investing in quality, the ones who refuse to settle for AI slop, you're going to be the ones that shape what this era looks like.

Designed to be read: making a system AI can actually use ft. Harvey Whiting (Meta) | Config 2026

Scripts