Every time I remade a company or service deck, the thing that drained my hours was never the content — it was reproducing the brand by hand. Copy a template, place the headings, set the brand colors, lay out the diagrams, fix the spacing. Add one page and its corner radius and line height drift slightly from the page before. Half a day quietly dissolves into detail work.
I've built smartphone apps solo since 2014 as an indie developer, and I now run four technical blogs in parallel, including Claude Lab. I make a lot of decks, yet rebuilding the brand by hand each time was so heavy that a single polished deck took days. That dissolving time mostly disappeared once I started using Claude Design. Here is how it went, including where it tripped me up.
If you need the basics — how to access the research preview and which prompts land — I covered those in Getting started with Claude Design. This article focuses on turning existing material into a design system and producing decks from it.
Why "the system stays" beats "it comes out fast"
Claude Design is a visual tool Anthropic Labs released on April 17, 2026, running on Claude Opus 4.7. Chat on the left, canvas on the right: describe what you want, a design appears, and you refine it through conversation. It accepts text plus images, DOCX, PPTX, XLSX, and web captures, and exports to PDF, PPTX, Canva, HTML, and shareable links.
That all sounds like "a tool that produces pretty decks fast." But the value I actually felt wasn't speed. Once you teach it your brand as a design system, everything after that comes out in the same unified look, in a fraction of the time, no matter who makes it. The brand stays as a reusable foundation rather than a one-off output.
When you run apps for years, treating decks as disposable outputs wears you down. The real question is whether you can shift to a structure where the brand accumulates as an asset, and the second deck onward needs only its script. For an individual or a small team, that shift is where productivity is won or lost.
Feed it one existing slide to extract the design system
The first thing I did was hand it my existing company PPTX as-is and let it extract a design system. The steps are simple:
- Open Claude Design and create a new entry under "Design systems"
- Write a description of the brand
- Pass as many reference assets as you can — existing slides, logo, color palette, GitHub repo, Figma files
- Run the generation and wait
Step 3 is the point. Passing finished, real artifacts raises accuracy far more than a description does, because one completed page carries more information than paragraphs of specification. Anthropic frames it well: one finished landing page tells the model more than a color-palette spec sheet.
After a short wait, it had verbalized the type scale, brand colors, card variations, and a spacing scale — and even the tone of voice, like "use concrete numbers" and "no emoji." The brand character I could never quite articulate myself had become reproducible rules. That took about fifteen minutes.
Pass a script and watch the design system pay off
With the system in place, the rest is content. Under "Slide deck" I selected that design system, created a project, and pasted the deck outline into chat. From the first frame, the slides respected the brand color, the title bar, and the per-service icons. Even the painful "role × phase" matrix table took shape through dialogue. STEP 1 and this together took under an hour — against days before, the difference in feel is night and day.
What I couldn't hand off — spacing, "too literal," complex diagrams
Honestly, it wasn't done by dropping everything on the AI. Three things needed my hands. Spacing and layout were 90% pleasant out of the box, but the last 10% was faster to fix with the "Edit" panel by setting px and color directly than by asking in chat. A script mapped one-to-one onto slides gets either too dense or too sparse, so I planned on a human pass. And complex diagrams: as a non-designer, I lacked the shared vocabulary to describe a layout in words alone, so the "tighten this, align that column" round trips piled up.
For complex diagrams, draw a rough first and hand it over
What fixed the diagram problem was sketching a rough first with ChatGPT image generation, then handing that to Claude Design with "make the same diagram." Showing a picture beats describing composition in words, and redrawing a single image is cheaper in tokens, so rough via image generation → clean version in Claude Design was efficient on both time and cost.
This also clarified a line I keep. I'm an artist — my photo-collage work has been recognized in international design awards — yet I use no AI in the artwork itself. For operational, outward-facing design work like decks, app icons, and store images, I lean on AI fully. That separation is how I protect time for the actual creative work.
If you want to try it, start by handing over the single best deck you already have and extracting a design system from it. That one page describes your brand more accurately than any spec you could write. Thanks for reading.