Facebook's Release Just Gave Claude Code A New Design System

AAI LABS
Computing/SoftwareSmall Business/StartupsInternet Technology

Transcript

00:00:00Now Meta just did something huge. They released their own design system. And this isn't just any
00:00:04design system because it has two things you won't find in anything else. The first is that this is
00:00:09Meta's own system, the exact one their engineers have built and refined over eight years and still
00:00:14run across their apps. So it comes already proven at scale instead of being some fresh experiment.
00:00:19The second is the bigger one. It's the first design system actually built for AI agents to
00:00:23read and build with, not just for humans. Everything about it is set up so an agent can
00:00:28understand the whole system and use it directly instead of guessing its way through like it
00:00:32usually does. Now if you don't know us, we're a software company and this is our channel AI Labs
00:00:37where we show you how to optimize processes with AI just like how we've optimized our own. And in this
00:00:42video, we're going to go over Meta's new release in detail, how it actually is agents first and the
00:00:47skill we built to get far more out of it than it gives you on its own. Before we get into how this
00:00:51system looks, let's quickly cover what Asterix actually is. So Asterix is basically Meta's own
00:00:57internal system for building sites and they've just open sourced it. Meta's style is professional and
00:01:02minimal. It's the same design approach that shows up across Meta's own products. So minimalism mixed
00:01:08with brand level design and systems thinking. That way, every part of a site follows the same rules
00:01:13and fits together. It's fully customizable, so anyone using it can tweak it to their own brand colors and
00:01:19make Meta's system their own. It's built on React and Stylex, which are basically two of the most popular
00:01:24tools people use to build the front end of a site. And it's a component library kind of like ShadCN,
00:01:30where ShadCN gives us pre-built components we can drop straight into our app. Asterix gives us over
00:01:35150 Meta-style components we can tweak to what we need. It's also designed so agents can use it easily.
00:01:41It ships with ready-to-use templates for the pages that come up the most, like the login, settings,
00:01:47and chat pages, and the agent can pull those in directly as a starting point. You can even use the
00:01:51playground to see how your changes look. It shows the code on one side in an editable pane, and any
00:01:57change you make there shows up right away in the pane on the right. The playground's also where you can
00:02:01explore Asterix's themes. Each one has its own style, colors, and fonts that suit a different kind
00:02:07of website, and you can preview them right there to see which one you want without actually applying it
00:02:11to your project. Since it's still in beta, the engineers at Meta are actively updating and improving
00:02:17it. They built it off their 8 years of experience designing systems at the company, and it's already
00:02:22powering a huge number of the apps they've shipped. The main principle behind the library is guidance
00:02:27over enforcement, so instead of adding guardrails that stop agents from doing certain things, it just
00:02:32gives them components they can use directly. The guidance doesn't live in prompts, it lives in the docs and
00:02:37examples, because every convention needed to build with Asterix is documented, down to each rule and
00:02:42detail. So Asterix is built so both humans and AI agents can work with it without getting stuck. And
00:02:47because meta engineers build and maintain it, everything in the design system is tested and
00:02:52proven in practice, instead of based on assumptions about how things should work. Now that we know what
00:02:57Asterix is, the next step is getting it into a project. The install's pretty straightforward. You'll find
00:03:02the steps on GitHub, where you copy the install command for the package manager you're using.
00:03:07A package manager is basically the tool that installs and manages your project's code libraries.
00:03:12We were building our project with npm, so we copied that command, opened our project folder and pasted
00:03:17it in. Once that command ran, it installed both the core and the theme of Asterix. You can swap this
00:03:23theme for any of the ones we saw earlier, and since we liked the neutral theme for what we were building,
00:03:27that's the one we installed. Aside from the theme, the command also installs the CLI and adds
00:03:32Asterix as a project dependency, so the library's components can be used in the app. The CLI acts
00:03:38as a bridge between the agent and Asterix, and it lets the agent access and download the components
00:03:43it needs to build the site. Asterix also has an MCP, but the CLI is the better option because it only
00:03:49loads what it needs bit by bit and doesn't fill up the context window. There's one more thing you need
00:03:54to run if you're building with agents the way we always do with clod code, and that's the Asterix
00:03:58init command. What this init command does is get the project ready for AI coding agents to use. When you
00:04:04run it, it'll ask you a few questions, and it'll also ask whether you want a template. But if you're
00:04:09building on Next.js like us, there's one thing to watch out for. The template it installs lives in
00:04:13the pages folder, while Next.js uses an app folder. All the new Next.js projects use the app folder for
00:04:19creating pages, which is called the app router. The pages folder on the other hand means the older
00:04:24page router is being used. And the app folder and the pages folder can't both exist in the same
00:04:29project, having both will throw an error. So you need to move that template into the app folder,
00:04:34or change the page name to app during installation, so this doesn't cause conflicts in future runs.
00:04:40The init command changes the clod.md and agent.md files, which are basically the guiding files for
00:04:46whatever agent you're using to build. It adds all the instructions that help the agent know
00:04:50how to build with the Asterix CLI, which commands to run to get components and what process to follow.
00:04:56With those instructions, the agent builds the app properly using the CLI underneath,
00:05:01instead of guessing on its own or reaching for web search. And that matters because an agent working
00:05:05on its own doesn't quite get things right, as you might have experienced yourself. Asterix fixes this
00:05:11by keeping the agent grounded in documentation, so it works from documented facts instead of guesses.
00:05:16Through the clod.md file, the agents prompted to look for those docs using the CLI. The same prompt
00:05:22goes into the agent.md file too, so even if you use some other agent instead of clod, it still gets
00:05:28pointed to them. Even though we picked the blank template, its styles still fill up the globals.css
00:05:34file, which holds all the design and theme styles that apply across the whole website, so they can be used
00:05:39anywhere. The page itself renders empty for now, just a blank page, but from here we can use those styles
00:05:45to build out our site. And if you're enjoying the video so far, subscribe to the channel and hit the
00:05:50hype button. This small gesture of support goes a long way for us. With Asterix installed and the
00:05:55project ready for agents, we can start building an actual site. To use it, you just give the agent a
00:06:00prompt describing the website you want in detail. Once you send it, Claude starts by orienting itself with
00:06:06the project first, since we'd already installed the Next.js project. It then follows the Asterix workflow
00:06:11spelled out in the Claude.md file. Claude then calls the Asterix CLI tools and looks for the starter
00:06:17pack, the design prompts, and any other resources Asterix has that match the design direction we're
00:06:22going for. That way it doesn't rebuild things from scratch, it starts from what already exists.
00:06:27From there, Claude picks the template that fits best. It goes through multiple commands to pull all the
00:06:32components it needs. Then it starts writing the code for the app and runs it. The site it built was simple
00:06:37and did follow the minimalist style, though there were a few issues. One card was wrapping to a new line
00:06:43when it should have stayed aligned with the main line and the products weren't centered. The color theme
00:06:47it picked also looked like AI slop that usual blue and black theme AI models reach for while building
00:06:53almost any business website. And even though we were only building a simple site, it took a really long time,
00:06:58around 16 minutes, way longer than we expected. But before we move forward, let's have a word by our
00:07:03sponsor, Lychee Gen. I took this character artwork and a couple of minutes later, it was a fully 3D,
00:07:09ready to print model, no modeling skills needed. Here's how. With Lychee Gen, you go from an image,
00:07:14a sketch, or even a text prompt to a 3D model in minutes. I uploaded my artwork and it generated a clean,
00:07:21printable model right on screen. My figure looked great, but the sword felt too small. So in the edit
00:07:26studio I just said, make it bigger, without starting over. Then I opened it straight in Lychee Slicer,
00:07:31where it came in clean, watertight, and ready to print. And here's what actually matters. Lychee Gen
00:07:36is built by the team behind Lychee Slicer, the most popular resin printing software with over 800,000
00:07:42users. So these models aren't just pretty on screen, they're built to print, with clean topology and
00:07:48auto repair baked in. Every new account gets free credits, so your first model is completely free.
00:07:53Click the link in the pinned comment to get started with Lychee Gen.
00:07:57To fix that generated look, we first had to figure out exactly what was causing it. If you've watched
00:08:02our videos, you might remember the AI Labs design system we've talked about a lot before. For those
00:08:06who don't know, it's basically a collection of all the skills we use across the projects we build,
00:08:11whether internal or external. These skills have genuinely helped us. Inside those skills,
00:08:16there's an AI Slop Detector skill. Its main job is to run sites built by agents through a critical
00:08:21review session that finds all the patterns that count as AI Slop. That way, the design we build
00:08:27doesn't look generated, it looks designed. The skill produces a report, but it doesn't fix the issues on
00:08:32its own, so we can iterate and fix them ourselves. It's constantly evolving and already catches a lot
00:08:37of known patterns, and if you tell it something still looks like AI Slop, it adds that to its patterns,
00:08:42too. So we pulled that skill and ran it against the website. It started auditing the landing page,
00:08:47ran the scanner, and found all the issues. It reported that while the visual design was clean,
00:08:51it looked like a textbook-generated SaaS landing page. The scan came back with four findings that
00:08:56made it look like an AI Slop site, and the highest priority one was the overall look of the landing page.
00:09:02We then prompted it to run fixes for those issues. Once the skill ran the fixes, it changed the color
00:09:07palette from that typical AI Slop 1 to something more minimal. It fixed all the reported
00:09:12findings and cut the cliched wording so the site actually looked designed instead of generated.
00:09:17The result was way better than the first version. The next step was getting Asterix to build an
00:09:22impressive site on its own instead of a basic one we had to fix afterward. To get more out of the
00:09:27library and fix all the issues we were running into earlier, we made a skill called Asterix Max.
00:09:33The way this skill works is, instead of following the generic layout we saw the library use, where it
00:09:37leans on bare templates and builds simple sites, the skill pushes Asterix to its highest limits.
00:09:43The most important thing it does is ground the model. Basically keep it working from real facts
00:09:48instead of guesses. The way it does that is by running the Asterix manifest command, which creates
00:09:52a file that lists everything available. That makes the model aware of all the commands and flags it would
00:09:58otherwise miss, simply because it didn't know they existed. With that manifest, the model doesn't have to
00:10:03guess. It knows what exists in the CLI and can fix the issues it runs into. The skill then follows the
00:10:09instructions in the claw.md file and uses tools to actually see the design, checking things like clean
00:10:15grids and proper depth hierarchy, so those issues we hit don't happen again. It also has references across
00:10:20multiple areas and it includes evals, basically automatic checks to spot and run tests for issues. To test
00:10:27the skill we'd built, we also installed the AI Slop Detector skill here, so the two could work
00:10:32together to generate a better site. That meant we wouldn't have to fix it later. Everything could
00:10:36be done in a single run. We called the skill and told it the site we wanted to build. Like we instructed,
00:10:41it ran the manifest command first and gathered all the content it needed. It then followed a workflow
00:10:46similar to the one it used before to build the site. After that, the skill started using a headless
00:10:51chrome browser, basically a browser that runs without a window you can see, to check how the site looked.
00:10:56Working together with the AI Slop Detector skill, it repeatedly took screenshots,
00:11:00analyzed them and fixed the issues it found. By the end, all the changes were in place across
00:11:05the site. Even though this took longer, because it was reviewing and improving at the same time,
00:11:10the site that came out was significantly better. With this complementary skill, there were still a
00:11:14few issues, but we were able to fix them pretty effectively. The site was far better than Asterix
00:11:19working on its own. Now, the skill we just built, along with the AI Labs design system, can be found in
00:11:24AI Labs Pro, which is our community. That's where you'll get all the resources, the starter packs and
00:11:29more, along with a place to interact with a bunch of like-minded people, including our team. So if you've
00:11:34found value in what we do and want to support the channel, this is the best way to do it. The link's in
00:11:39the description. That brings us to the end of this video. If you'd like to support the channel and help
00:11:43us keep making videos like this, you can do so by using the super thanks button below. As always,
00:11:49thank you for watching, and I'll see you in the next one.

Key Takeaway

Integrating Meta's Asterix design system with custom agent-based workflows like 'Asterix Max' and automated visual scanners enables the creation of high-quality websites while preventing generic, AI-generated design patterns.

Highlights

  • Meta's open-source design system, Asterix, contains over 150 component libraries built on React and Stylex.

  • Asterix is engineered for AI agents to read and utilize directly, moving away from human-only documentation.

  • Using the Asterix CLI over an MCP allows agents to load components incrementally, preserving the LLM context window.

  • The Asterix Max skill automates site construction by running a manifest command, allowing the AI to access specific CLI commands and flags.

  • Pairing the AI Slop Detector with automated browser-based visual reviews significantly improves the quality of AI-generated websites.

Timeline

Meta's Agent-First Design System

  • Meta released Asterix, a design system based on eight years of internal development.
  • The system provides over 150 components specifically structured for AI agent readability and implementation.
  • Asterix prioritizes guidance through documentation and examples rather than enforcing rigid constraints.

Asterix is an open-source design system built on React and Stylex that brings Meta's internal minimalist design aesthetic to external projects. Unlike human-centric systems, it is explicitly built so agents can understand and use the components directly. It provides ready-to-use templates for common pages like logins and settings.

Implementation and Agent Integration

  • The Asterix CLI connects AI agents to the component library.
  • Running the 'init' command configures agent guiding files like Claude.md and agent.md to prevent AI from guessing.
  • Users must ensure template paths align with the Next.js app router to avoid folder conflicts.

Setting up the project involves using a package manager to install core libraries, themes, and the CLI. The CLI serves as a bridge, allowing agents to download only necessary components. Modifying the agent instruction files forces the model to work from documented facts rather than reaching for web search or unguided assumptions.

Optimizing AI Output

  • Default AI site generation often produces 'AI slop' characterized by generic color themes and poor layout alignment.
  • The AI Slop Detector skill audits landing pages against known patterns to identify and correct generic design issues.
  • Asterix Max, a custom skill, uses the manifest command to ensure the agent understands all available CLI tools and flags.

Standard AI site builds using Asterix can result in lengthy build times and generic outputs. By implementing an auditing skill that scans for AI-specific design patterns, users can automatically identify and fix issues. Further performance gains occur by forcing the agent to use a manifest file and headless Chrome for real-time visual review during the generation process.

Community Posts

View all posts