Claude Design is Anthropic’s design tool, at claude.ai/design. Where a plain chat answers a design request instantly, this one interviews you first, keeps your brand’s colors and fonts on file, and lets anyone on the team fix a draft by circling the problem. Austin Marchese, a creator who uses Claude Code daily, walked through seventeen uses of it in an Anthropic-sponsored video. This guide distills them into one workflow.
01Trained to ask design questions
Claude comes in several general-purpose products (the chat, Claude Code for coding, Cowork for office work) and this specialized one. The difference shows the moment you ask for something visual. Marchese asks for a party flyer, and instead of producing one, Claude Design asks who is hosting and what paper the flyer will print on. The same prompt in a plain chat generates an image immediately and never asks.
“It'll ask you questions that normal Claude just would never think to ask.”
Austin Marchese, “17 Things You Didn't Know Claude Design Could Do” · 1:06If an asset deserves a designer’s questions, it belongs here; for a quick answer or working code, stay in a plain chat or Claude Code.
One limit to know going in: it is a layout tool, in the same category as Canva or Figma, and it will not generate photos or images. Where a picture belongs, it leaves a labeled empty box, a placeholder, for you to fill. Generate the image elsewhere and drop it in; the design and the layout are what this tool decides.
02Start with a design system
The usual failure with AI-made assets is that every one starts from scratch: you argue a report into the right look, and the next report forgets everything. The fix is a design system, the written record of your brand’s colors, fonts, and layout rules that every later asset inherits.
In Claude Design that record is a built-in feature. Open the design system tab, attach your website so it can read your existing branding, answer its questions (or click “decide for me”), and it produces a system the whole organization can use. It is the same move One skills library for your whole team makes with written procedures: build the standard once, and everyone’s output inherits it.
Once the system exists, it feeds four kinds of work:
- Rebuild old documents. Marchese uploads a client proposal made in Canva under his old branding and has it rebuilt in the new system, then tweaks it directly on the page. Any stale deck, one-pager, or PDF gets the same treatment.
- Rebuild in the client’s branding. The same trick reversed: hand a prospect a proposal that looks like their own company made it. Give the tool your proposal and the client’s site, and ask for the remake in their style.
- Templates for the rest of the team. A designer saves an Instagram graphic, weekly report, or proposal as a template inside the design system, and non-designers fill it in without breaking the brand.
- Working email templates. He shows it a plain login email, asks how to move the redesign into Supabase (the service that sends his login emails), and it rewrites the email with the per-user parts, like the login link, marked the way Supabase expects, ready to paste in.
03Fan out variations, then pick
Describing a design change in words is slow and imprecise. Marchese’s habit is to make the tool show him options instead: ask for five variations of a landing page, or of one piece of it like a header or a chart, pick the two he likes, ask for five more in that direction, and converge over a few rounds. Five is a starting point, not a rule.
Color gets the same treatment. Ask it to explore color and type pairings for your channel or product, and it renders them on your actual design; a built-in switcher lets you click between pairings and watch the page change. Seeing options beats reasoning about them in the abstract, and clicking between rendered versions beats editing code and reloading.
The tool can also decide how an asset gets used. He has it model a 3D version of his logo, likes it, and has no idea what to do with it, so he asks the tool to put it on his website in a few realistic ways. It returns three options and he picks the slowly rotating version. When you have an asset and no plan, ask the tool how it would use it. What makes an option good rather than merely different is its own craft; The design concept does the work covers where design quality comes from.
04Wireframe before you build
A wireframe is a clickable mock-up: it looks and feels like the product, but nothing behind the buttons is real. At Marchese’s last startup, pitches to artists like Ed Sheeran were wireframes, because clicking through a fake product persuades better than pictures of one, at a fraction of the engineering cost.
Claude Design makes them from a prompt. He describes a workout-plan product, lets the tool decide the open questions, and gets screens he can click through to feel how the app would work. Coding agents have made real products cheaper to build, but building one is still the expensive way to test an idea; a wireframe is often enough. When a design settles, connect your code base and have the tool port it over as the starting point for the real build.
05Comments the AI acts on
Feedback on a design usually degrades into words about pictures: screenshots, arrows, “a bit higher”. Claude Design keeps the feedback on the design itself. Marchese reviews a website draft, adds a comment telling the header to span two lines, draws a circle around a graphic where he wants it moved, then clicks “send to Claude”, and the tool works through the comments and fixes them.
The same mechanism carries a team. Share the design and teammates comment directly on it; nobody installs anything, pulls the latest code, or trades screenshots. A reviewer with no technical background can mark up the draft the way they would mark up paper, and the AI applies the changes.
06Edit it until it reads like you
Everyone has a backlog of assets that never get made because they are nobody’s job: the how-to video you wish existed, the onboarding explainer for a new teammate, the one-page PDF version of something you published, the resume untouched for eight years. Marchese makes several on camera: an animated explainer of his product, a branded PDF companion to a video, a resume rebuilt from his LinkedIn. His rule of thumb: use the generated version where the stakes are low and the asset would otherwise not exist; a product launch still deserves exact, by-hand control.
The reason generated assets are often embarrassing to send has a name in his telling: doom prompting. You go back and forth with a chat, it never quite lands, you give up and export the near-miss anyway.
“AI slop is destroying what people think about what you produce.”
Austin Marchese, “17 Things You Didn't Know Claude Design Could Do” · 16:15Claude Design’s answer is editing in place: click into the draft and rewrite the text directly on the page until it sounds like you, no more prompting round-trips. The design carries your brand because of the system; the words carry your voice because the final pass over them was yours. For why generated output drifts into slop in the first place, and how to catch it in text too, see Why AI makes slop and how to catch it.
Further reading
- Austin Marchese · 17 Things You Didn’t Know Claude Design Could Do, the video this guide distills
- Anthropic · Claude Design, the tool itself
- Anthropic · Claude product page, where Design sits among the Claude surfaces