I Gave Claude a Brand. It Gave Me Back a Color Palette I Couldn't Edit.

Anthropic's new design capabilities in Claude are genuinely impressive. The model can reason about layout, generate working interfaces, and produce polished visual artifacts in a single turn. As someone who builds AI tools for presentations, I had to try it for the use case I care about most: can it actually make a deck?

I gave it a real test. Not a toy prompt. We created a fake brand for this exercise, Heave, a ride-sharing company, and built out the kind of materials any real company would have. We uploaded a fully branded PPTX deck with finished example slides, gave Claude a functioning website for the company, and shared a brand kit explaining the ethos, color system, typography, logo usage, and illustration style.

A coherent system. A single dominant green, one signature font treatment, custom illustrations with a distinctive style, and a confident slide rhythm, exactly the kind of brand most enterprise customers walk in the door with.

What Claude Produced

Claude built me a single-page HTML document.

Not a PowerPoint file. Not a Google Slides deck. Not anything I could open in Keynote, edit before a meeting, or send to an investor who expects a normal attachment. An HTML page with sections styled to look like slides, a hero block, some feature cards, a metrics row, a closing CTA, rendered as one long scrollable web page.

At first glance the output is convincing. The green is right. The layout instincts are sound. The numbers are organized cleanly. Squint and it almost looks like a real Heave deck.

Then the cracks show. The typography is wrong, Claude defaulted to a serif display face that isn't in the brand kit, and the body type doesn't match the system font Heave actually ships with. The custom isometric illustrations and brand photography I uploaded? Gone. The logo? Never appears once. Every slide is text, numbers, and a single accent color. The distinct visual personality of the brand has been flattened into a competent-but-anonymous editorial template.

And because the artifact is HTML, none of it is editable in the way a real deck is. I can't drag a number, swap a chart color, replace a metric, or fix the typography myself. Every change has to go back through chat: "make the headline smaller," "use sans-serif," "add a logo." Each tweak is a new generation, more tokens, and another round-trip, versus opening the file and clicking the text.

To be fair: it looked nice. Clean typography choices in isolation, sensible hierarchy, decent spacing. As a landing page concept, it was solid. As a presentation, it was unusable.

The Three Things That Stood Out

1. The output format is wrong for the job

Presentations live in PowerPoint and Google Slides for a reason. Sales teams send .pptx attachments. Customer success teams paste decks into QBR Zooms. Boards review printed slides. Marketing teams hand off final files to event AV crews. The entire enterprise workflow assumes a real, editable, shareable presentation file.

An HTML artifact can't enter that workflow. You can't open it in Slides to tweak a number before a meeting. You can't email it to a CFO. You can't drop it into a deal room. It's a beautiful demo of what AI can render, not a deliverable that fits how teams actually work.

And getting anything closer to a real file turned out to be impossible. There's no native "export to PowerPoint" button on the design output, so I had to chat-request a PPTX directly from Claude. The model invoked an export skill, started converting, and then the wheels came off. It asked me to confirm which tab was open in my preview pane, whether I could "see the deck," to click specific HTML files in my tab bar, to refocus the preview, to manually refresh it. Each retry hit the same wall. After several rounds it offered me a PDF instead "which uses a different mechanism that may work more reliably."

I never got an editable file out of it. Not once. The closest I came was a static PDF, which is exactly the same problem as the HTML artifact: a thing you can look at, not a thing you can edit, version, or hand to a teammate.

This is the gap tools like Gamma have been trying to close for years (with their own proprietary format), and that Decky closes by going the other direction entirely: producing native Google Slides and PowerPoint files that drop straight into the existing workflow.

2. It picked up the colors, and nothing else

This was the most revealing part. Claude did notice the brand. It pulled the Heave green and built a deck where the color palette was broadly correct. On a quick glance, it looked like a Heave deck.

Look any closer and the illusion falls apart. The custom isometric illustrations, the single most distinctive part of the Heave identity, were nowhere in the output. The logo never appeared. The signature font treatment from the example slides wasn't there. Photography from the brand kit was ignored. Claude essentially threw away every visual asset I gave it and substituted generic, AI-generated equivalents tinted in the brand color.

But here's the thing: this is what every generative AI presentation tool already does. Gamma, Beautiful.ai, Tome, Canva's AI, they all let you pick a color palette and apply it to a generic template. That's been the default for years. It's a theme picker dressed up as brand awareness.

A real brand is not a color. It's a system, typography, photography, illustration style, layout grids, the way a chart looks, the way a logo is locked up, the rules for what not to do. When an AI tool tells you it "matches your brand" but is really just matching your hex codes, the output is usable, it just looks like every other company that picked a similar palette. That is the definition of brand drift, and at enterprise scale it's the entire problem worth solving.

3. Token usage was high, really high

I didn't instrument this scientifically, but I watched the token meter climb. A single 10-slide attempt burned through what felt like a meaningful chunk of a Pro-tier monthly allowance. Multiply that by an enterprise where hundreds of employees might generate dozens of decks a week, and you're looking at a serious cost question, even before you factor in that the output isn't usable.

The economics of "regenerate until it's right" don't work when each regeneration costs real money and the underlying format problem can't be regenerated away.

What This Tells Us About AI Presentation Tools

Claude's design capability is a genuine technical achievement. It also clarifies, by contrast, what the next phase of AI presentation tools actually has to solve.

Generation isn't the hard part anymore. Any modern model can produce slide-shaped content. GPT, Claude, Gemini, they can all write headlines, structure narratives, and lay out information. That problem is solved.

Fitting into the workflow is the hard part. Enterprise teams don't need a new artifact format. They need on-brand decks that open in the tools they already use, look identical to what their design team would build, and respect the brand system that took years to develop.

Color matching is not brand matching. When a tool picks up your hex codes and calls it "on brand," it's doing what every theme picker has done for a decade. Real brand fidelity means using your actual typography, illustrations, photography, and layout system, not approximating a vibe with the right green.

The Honest Take

Claude Design is impressive technology, and I'd encourage anyone building in this space to spend time with it. But "impressive demo" and "production tool for a brand-conscious team" are different things, and the gap between them is exactly where the real product work lives.

For an individual making a personal pitch or a one-off concept, Claude can probably produce something good enough. For an enterprise team that needs every deck to be on-brand, editable in Google Slides or PowerPoint, and consistent across hundreds of users, that's a different category of tool, and it requires a different architecture from the ground up.

That's the bet we made when we built Decky: that the right output is a real Google Slides or PowerPoint file, the right brand system is your existing assets (not a theme picker), and the right place for the AI to live is inside the workflow you already use, not in a new browser tab that produces an HTML artifact you can't actually deliver.

Claude is going to keep getting better. Everyone in this space is. The interesting question isn't whether AI can design a slide. It's whose slide it ends up looking like, and where you can actually use it.