After the release of Opus 4.5, Anthropic product designer Nate Parrott found himself overwhelmed by the surge in engineering output and used his spare time to develop Claude Design. He initially tried having AI generate designs directly, then shifted to leveraging AI’s strength in writing HTML, using HTML as a unified visual canvas. He also distilled Anthropic’s brand system into prompts, ensuring that generated designs naturally adhered to brand guidelines. This side project was widely adopted by colleagues internally for creating presentations and was eventually formally adopted by the Labs team, launching as a research preview in January 2026. Claude Design is positioned as a pre-production tool, excelling at generating interactive prototypes, presentations, PDFs, and other visual communication assets, while leaving actual production development to Claude Code. As the model takes on more coding tasks, the designer’s value shifts toward the front end—from “producing visuals” to defining direction, building systems, and driving alignment. Within its first week online, the product was used by over one million people.

Author and source: AI World

Engineering takes off first, while design is left behind.

Claude Design turned out to be a side project that emerged unexpectedly.

Behind it is a story of two engineers and one designer.

The story begins with the release of Opus 4.5 in November 2025.

At that time, Nate Parrott was the only designer on the Claude Code VS Code project, working alongside two engineers.

Anthropic product designer Nate Parrott recounts the entire creation process of Claude Design in a first-person account on the official blog.

The three of them worked together to transform Claude Code, which could only be used by typing commands in a terminal, into an interface that anyone can understand and use easily.

At the end of September, the product entered beta and everything was normal; then in November, Opus 4.5 was released. Here’s how Parrott described what happened next:

Two engineers suddenly began delivering at a frantic pace, producing far more than before, while he remained at his usual speed.

Soon, he fell behind.

“I need to find a way to catch up,” Parrott thought.

So he spent his free time building a small tool specifically designed to help him keep up with the project's pace.

And this "side project" became Claude Design, which was used by over one million people in its first week online.

Engineering takes off first, while design is left behind.

The heartbreaking part of this story is that AI didn't raise everyone's speed equally—it first accelerated engineering.

As a result, the design by Parrott quickly became the slowest link in the entire production line.

The turning point is November 24, 2025, the day Opus 4.5 is released.

According to Parrott’s recollection, since then, the entire Claude Code team has significantly accelerated its delivery pace, with the two engineers producing much more than before, while his design output remained at the same speed.

This forces him to find a way to catch up.

This also reveals the most overlooked aspect of the AI wave: the initial impact is often not whether a particular job will disappear, but rather how it disrupts the established rhythm of alignment between workflows.

For example, once engineering accelerates, any team that can't keep up—design, product, marketing—will instantly become a new bottleneck.

One-month side project, from "AI writes design" to "AI writes HTML"

Parrott's original idea was straightforward.

Claude Code is a purely text-based terminal tool, and initially, it approached design as a textual task:

Package the terminal output, a few interface screenshots, and a single requirement, and tell Claude, “This is the feature we want to add—design it for us.”

However, the results were not ideal.

For the next month or so, he treated it as a side project—officially unapproved, but progressed gradually during his spare time.

Slowly, he stumbled upon the real solution: Claude is especially skilled at writing HTML.

We are accustomed to thinking of HTML as something used for creating web pages.

But Parrott’s epiphany was that HTML is actually a rich interactive visual medium: anything you can do with slideshows, video files, or PDFs, you can do with a single webpage.

So he stopped asking Claude to "draw designs" and instead had it "generate HTML," then wrapped it in an interface: chat on the left, real-time preview on the right. Say something, and the right side updates accordingly.

Parrott's breakthrough wasn't in "making AI draw images," but in treating HTML as a unified visual canvas: presentations, interactive prototypes, animations, and web pages can now all flow through the same generation and editing pipeline.

Claude Design’s interface: conversation on the left, canvas on the right—exactly as Parrott originally envisioned in his prototype.

HTML alone isn't enough—the brand system is the key.

Knowing only HTML is still far from being "usable."

In Parrott’s view, the core of product design is translating your understanding of the product and brand into every specific design decision.

So he spent considerable time distilling Anthropic’s brand elements—fonts, colors, assets, and design principles—into the prompts.

This way, when he inputs requirements into AI, the generated designs are inherently more aligned with Anthropic’s guidelines.

He fitted the entire set into a compact internal prototype and shared it with the team. Product designers immediately picked it up and used it to create interactive prototypes.

In traditional design tools, creating a clickable prototype means you have to draw every state of every screen and manually connect them one by one.

In Claude Design, you simply drop your assets into Claude and say, “Make it run”—it generates something with a built-in link that you can directly share with colleagues like you would send a document.

This feature later became the core capability of the entire product.

Today, when you start using Claude Design, it will scan your codebase and design files during the initial setup and automatically build a design system for your team.

Afterward, all projects will automatically apply your colors, fonts, and components. Large teams can assign an administrator to lock in a standardized system, preventing anyone from making unauthorized changes.

An offsite turned the side project into a full-fledged initiative.

What truly made Parrott feel “this thing was going to work” was a brainstorming session at Anthropic Labs where ideas were pitched.

He noticed that almost everyone in the room was using this tool to put together their slides, with many even creating theirs on the spot midway through the meeting, just before it was their turn to speak.

It was this moment that convinced the Labs team to officially assign personnel to the project. From this point on, a side project became a legitimate product.

Claude Design has already been tried by Anthropic once before.

On January 13, 2026, Anthropic announced the expansion of its Labs, bringing on Instagram co-founder Mike Krieger and Ben Mann to lead the initiative, with the goal of incubating experimental products pushing the frontiers of Claude’s capabilities.

Claude Code started as a research preview and grew into a $1 billion product in six months.

Claude Design followed the same path: starting as an individual’s private experiment, then gaining organic adoption by internal designers, followed by validation through peer support at idea pitch sessions, and ultimately leading to Labs allocating resources, releasing a research preview, and transitioning to beta.

No step in this chain was planned from the top down.

AI-native products are often not designed, but rather emerged from genuine internal use.

Value is shifted forward; designers are no longer just “producing visuals.”

After Claude Design was officially approved, its positioning changed quickly.

They no longer see it as just a tool for creating interface mockups.

It can handle anything used for visual communication—slides, landing pages, one-page printable PDFs, emails, animations, and social media graphics.

Parrott's own description is that it operates "one level above product design": what you refine with Claude are the visual assets primarily used for communication and ideation.

Roles are clearly divided: Claude Design handles the early stage, quickly exploring and comparing directions to build consensus; Claude Code handles the later stage, executing the actual production development.

The two can also be interconnected: pull in the design system using /design-sync, and make direct changes in the terminal using /design, with MCP serving as the bridge between them.

Claude Design can export to PDF, PPTX, and HTML, and integrates with tools such as Canva, Adobe, Gamma, Miro, Replit, Vercel, and Wix.

You can export your completed work as PDF, PPTX, or HTML, and one-click send it directly to tools like Canva, Adobe, Gamma, Miro, Replit, Vercel, or Wix to continue working.

However, Claude Design is not an image generation model. Parrott admitted: It’s not suitable for creating logos—you’ll need to upload your existing logos, images, and brand assets yourself.

For production software that is actually being deployed, Anthropic still recommends sticking with Claude Code.

Ultimately, Claude Design is a preliminary tool, not a mature production development platform.

The more the model can take over coding, the more valuable work shifts forward—designers' value increasingly lies in choosing directions, building systems, evaluating quality, and driving consensus.

Go back to the original scenario.

When two engineers around him suddenly sped ahead, Parrott, the only designer on the team, was left behind. Instead of waiting, he took matters into his own hands and built the tools he needed to catch up.

Who could have imagined that this tool would later be used by millions of people?

So, when you feel anxious about your job being disrupted by AI, why not follow Parrott’s example: use AI to create a new tool for yourself—or find one that can help you open up new possibilities.