Anthropic launches Claude Design research preview with codebase-derived design systems

• Anthropic shipped Claude Design as an Anthropic Labs research preview, and [claudeai's launch post](src:0|claudeai's launch post) says it is powered by Claude Opus 4.7 and rolling out to Pro, Max, Team, and Enterprise plans.
• According to Anthropic's help guide, Claude Design takes screenshots, documents, and linked codebases as context, while [claudeai's product thread](src:1|claudeai's product thread) says it can infer a team design system from code and design files.
• Export is the whole point here: Anthropic's help guide lists PDF, PPTX, Canva, standalone HTML, Claude Code, local coding agents, and Claude Code Web, which lines up with [claudeai's workflow demo](src:1|claudeai's workflow demo).
• Early users on the main HN thread liked the comment-driven iteration and Claude Code handoff, while [theo's usage report](src:35|theo's usage report) and Anthropic's pricing page surfaced a separate weekly usage meter and some ugly failure modes.
• The market read was immediate, but the cleaner signal is product adjacency: TechCrunch's report says Mike Krieger left Figma's board days before launch, while [elder_plinius's prompt leak](src:28|elder_plinius's prompt leak) shows Anthropic built Claude Design as a filesystem-based HTML artifact generator with explicit handoff and verification steps.

You can browse the official announcement, skim the getting started guide, and the design-system setup doc is more revealing than the marketing copy. The help center quietly adds standalone HTML export, local coding agent handoff, and the fact that Enterprise has Claude Design off by default. Then there is the HN thread, where the interesting bits are not "AI made a mockup" but comment-driven iteration, backend scaffolding, and complaints about token burn.

Anthropic is positioning Claude Design as a separate Labs surface for prototypes, slides, one-pagers, and other visual artifacts, not as another chat mode inside the main Claude UI. The official launch post says it is powered by Opus 4.7 and available in research preview for Pro, Max, Team, and Enterprise subscribers.

The first-party workflow has three steps:

• prompt for a first draft
• refine through chat, inline comments, direct edits, or Claude-generated sliders
• export or hand off the result

The help center adds two rollout caveats that matter more than the promo video: Claude Design is default off for Enterprise, and usage is metered separately from chat and Claude Code.

The most differentiated claim is not image generation. It is that Claude Design can absorb an existing product's visual language from source material and reuse it across new work.

According to Anthropic's setup guide, the system can extract:

• reusable components
• primary, secondary, and accent colors
• typography choices
• layout and spacing patterns

The same guide says those patterns can come from several sources:

• a codebase or component library
• screenshots, prototypes, and existing design files
• slide decks or PDFs that reflect the brand
• individual assets such as logos and color specs

That makes the product more code-adjacent than the launch framing suggests. [claudeai's third post](src:2|claudeai's third post) explicitly says Claude can read a codebase and design files, then build and apply a team design system automatically.

Anthropic's docs are unusually explicit about where Claude Design is supposed to land. The getting started guide lists these output paths:

• ZIP download
• PDF export
• PPTX export
• Canva export
• standalone HTML export
• handoff to Claude Code
• send to a local coding agent
• send to Claude Code Web

That last cluster is the engineer angle. In the main HN thread, one top commenter said Claude Design made it easy to explore alternatives, tweak details with comments, then export to Claude Code including the design system. Another HN comment claimed a two-line prompt could bolt on a backend by spinning up cURL-driven requests against AGENTS.md, which is a much more agentic story than "Claude made slides."

The launch thread also keeps leaning on the Canva path. [claudeai's workflow demo](src:1|claudeai's workflow demo) mentions export to Canva, PDF, and PPTX, while [athasdev's UI screenshot](src:13|athasdev's UI screenshot) shows a tweak-heavy interface rather than a static canvas tool.

The early hands-on reports cluster around the same tradeoff: strong first-pass range, messy reliability.

From tweets and discussion:

• [altryne's long test](src:40|altryne's long test) called out live feedback, cursor-aware "point and talk," design-system generation from a GitHub repo, and PPTX export, but also said the session hit quite a few bugs.
• [theo's first report](src:35|theo's first report) said one project lost its files after burning 10 percent of usage.
• In a follow-up, [theo's usage breakdown](src:36|theo's usage breakdown) said two threads on the $200 tier had already consumed about half of the allowance.
• Later, [theo's second post](src:37|theo's second post) said the tool could produce good results with effort, which is about as honest a research-preview review as it gets.
• On Hacker News, one commenter praised the comment function and handoff flow, while another complained about massive token usage, weak logo work, forgotten feedback, and a tendency to restart from scratch.

Anthropic's own docs line up with those complaints more than the announcement does. The help guide lists disappearing inline comments, save errors in compact view, lag on large repositories, and "chat upstream error" failures that may require a new chat tab.

The ambient reaction on X was "Anthropic is coming after Figma," but the evidence is narrower and more concrete.

TechCrunch's report says Mike Krieger, Anthropic's chief product officer, left Figma's board on April 14, the same day The Information reported Anthropic was preparing an AI design product. That timing is why so many reactions treated the launch as a competitive shot rather than just another Labs experiment.

The product overlap is real in three places:

• prompt-to-prototype generation
• collaborative iteration on a shared canvas
• export paths into presentation and implementation tools

The counterpoint showed up fast too. [skeptrune's post](src:17|skeptrune's post) argued there is still a big gap between generating a first version and polishing it, and said incumbent design environments remain valuable because they absorb the maintenance burden Anthropic may not want. Even [altryne's hands-on verdict](src:40|altryne's hands-on verdict) stopped short of calling it a Figma killer.

The stock-chart screenshots, like [Yuchenj_UW's post](src:8|Yuchenj_UW's post) and [bridgemindai's later chart](src:20|bridgemindai's later chart), captured the mood. They do not establish causality, but they do show how quickly the market mapped Claude Design onto the design-tool category.

The leak is the clearest look at how Anthropic actually thinks this product should behave.

The GitHub dump linked from [elder_plinius's thread](src:29|elder_plinius's GitHub link) describes Claude Design as an "expert designer" working inside a filesystem-based project and producing HTML artifacts. It instructs the model to ask clarifying questions, explore design resources, build folder structure, surface a file with done, verify it loads cleanly, then call fork_verifier_agent.

A few leaked instructions stand out because they match public behavior almost one-for-one:

• output is a single HTML document unless the task calls for another format
• designs should inherit visual context from codebases, UI kits, screenshots, or brand assets, not start from scratch
• multiple variations should be exposed as tweaks or slides, not separate one-off files
• slide decks, videos, and fixed-size content should use their own JS scaling and persistence
• completed artifacts can call Claude from inside shared HTML via a built-in helper

That makes Claude Design look less like a generic canvas app and more like a packaged design-to-code agent harness with opinionated artifact formats. The launch tweet says "hand off to Claude Code," but the leaked prompt suggests handoff and verification were in the product grammar from the start.