Anthropic continues to launch innovative AI offerings at a rapid pace, with its newest creation, Claude Design, standing out as particularly noteworthy. This tool also promises to carry a substantial cost.
Launched today in a research preview phase, Claude Design generates prototypes for websites, wireframes for applications, presentation slides, promotional content, and various digital assets solely based on user instructions. Outputs can be downloaded in formats like ZIP archives, PowerPoint files, HTML code, or PDFs, or integrated with tools such as Canva or Claude Code.
The system operates using any Claude model, though Anthropic suggests employing the most recent version, Opus 4.7, for optimal performance. While it may appear as a human-replacement AI, its full capabilities likely require human guidance to shine.
Although users can input a basic command and observe the outcome, as I did, superior and unique outputs emerge when providing an established code repository and visual elements. For those with an active web presence, uploading site code alongside sample images helps Claude align with the desired aesthetic.
To maximize value from Claude Design, supplying detailed specifications about the desired outcome is essential, where human expertise in design proves invaluable. Greater initial clarity reduces the need for subsequent revisions.
Returning to my trial of Claude Design, I must note that I lack professional design skills.
Accessing Claude Design during its preview via the web interface greets users with a tabbed conversation area for entering the starting instruction. Options encompass initiating a fresh prototype, building a presentation, or working from an existing or new template. Additional features allow entering a business name, linking a GitHub account, attaching a local directory, or importing elements like typefaces and brand icons.
Opting for a minimal setup, I began without prior materials and entered a straightforward request: 'Develop an engaging visual aid that demystifies AI tokens for general audiences.'
Claude processed the input briefly before posing a series of selection-based queries: the intended viewers (options included complete novices, interested non-experts, and learners)? Preferred output style? Desired interactive features? Visual tone (e.g., professional like The New York Times or playful animation)? Scope of the content?
Responding to these took around a minute, after which Claude outlined its strategy—a sleek, informative piece with a 'New York Times or Pudding-inspired vibe, featuring serif titles, ample spacing, and a single highlight hue'—established five development stages, and proceeded to build.
Claude might pose clarifying questions following an initial Claude Design request.
As the creation unfolded, my dialogue with Claude shifted to a sidebar on the left, while the right portion of the screen became an expansive workspace displaying the evolving project. Upper tabs enabled switching between different iterations (Claude Design supports generating several simultaneously) and examining the underlying files.
In under five minutes, a preliminary version appeared—an attractive web page guiding users through AI tokens progressively, with hands-on areas for inputting text to observe live token tallying. The text was straightforward, approachable, concise, and, to my knowledge, correct. It closely matched my vision from the outset.
In total, Claude Design produced three variants of the AI token demonstration prototype in about 25 minutes. Quite remarkable.
However, checking my usage indicator revealed a surprise. Subscribed to the Claude Pro tier—intended primarily for personal, routine applications rather than enterprise prototyping—I had consumed 80% of my weekly Claude Design allocation. Currently, this usage tracks independently from the general Claude limits.
Claude generated this engaging interactive component to complement the token explanation page.
A confusion arose, compounded by my error. Attempting to navigate between prototype options and encountering a text alert for a 'preview token needed' (evidently, the variants weren't unified in one HTML document), I accidentally selected the 'undo' instead of returning, erasing all progress. 'The undo action removed everything. Reconstructing the files is necessary from the beginning,' Claude responded evenly.
With my usage bar nearly depleted, I opted for the more economical Sonnet 4.6 model as Claude addressed the issue. Yet, within another five minutes, my Claude Design weekly limit reached zero.
Fortunately, Anthropic had recently distributed extra credits to Claude subscribers (after restricting direct access to OpenClaw and similar external AI services without the Claude API for paid users), enabling me to replenish and complete the project. Even without my file-erasing mishap, I would have exhausted the quota midway through evaluation.
The end result appeared remarkably refined, considering my inexperience in design and the vague guidance provided. Nonetheless, this represents yet another resource-intensive Claude feature, one that Claude Pro subscribers can scarcely explore before hitting their caps.
Ben has covered consumer tech for over two decades, now specializing in AI's impact on daily life. His reporting examines cutting-edge language models and their practical applications in professional and personal settings to equip people for the coming AI era. 'AI will transform our world faster than anticipated,' Ben observes. 'Daily engagement is the key to adjustment.' A PCWorld contributor since 2014, he has reported on devices from computers to surveillance gear prior to spearheading the site's AI coverage. His work has featured in outlets including PC Magazine, TIME, Wired, CNET, Men's Fitness, and Mobile Magazine. Ben earned a master's in English literature.