view project

Meta data

  • Role: Senior product designer
  • Period: Feb 2025 – Jun 2026
  • Platforms: web app, website
  • Focus: end-to-end product: problem, research, UX, interface, AI interaction, production code, growth
  • Stack: Figma, AI coding (Claude), Webflow, Eleventy

Prezi AI

Prezi rebuilt its product AI-first. Slides are generated and edited through chat, and the presentation still moves with the animated camera Prezi is known for. I designed the core experience of that product and the surfaces that make it a business: sign-up, paywalls, the AI credit system, and the live engagement features that give people a reason to pay.

Ask Claude about this case study Ask ChatGPT about this case study Markdown

Prezi has been around since 2009. Its signature product runs on proprietary rendering technology, and years of accumulated decisions made every change slow.

When AI arrived, that stack became the bottleneck. Integrating LLMs into a proprietary format meant fighting the engine, and fast prototyping was not possible. LLMs are also simply better at HTML. It is the format they know best.

So the company started a new app. Slides are generated as HTML through chat, and the presentation keeps the thing Prezi is famous for: the animated camera that moves through a scene instead of flipping pages. The best of both.

The system has four layers, and each one only references the layer below it. Primitive tokens hold the raw values: the color scales, type sizes, and spacing steps. Semantic tokens give those values a job, like "brand background" or "error border," so designers and engineers choose by meaning, not by hex code. Components use only semantic tokens, and patterns combine components into reusable layouts. That indirection is what makes the system maintainable: change one primitive and every surface updates, or remap the semantic layer and the whole product switches to dark mode.

Diagram of the four token layers, each built only from the one before it. Primitive tokens hold raw values, such as blue/700 equal to #0F3ADD. Semantic tokens give a value a job: background/bg-brand-solid points to blue/700. Components use only semantic tokens: the primary button background is bg-brand-solid. Patterns combine components into layouts, such as a pricing card or a paywall. For dark mode, only the semantic layer remaps to different primitives, and components stay the same.

  1. Primitive Raw values blue/700 = #0F3ADD
  2. Semantic Values with a job background/bg-brand-solid → blue/700
  3. Dark mode The semantic layer remaps to different primitives. Components don't change.

  4. Component Built only from semantic tokens primary button background → bg-brand-solid
  5. Pattern Components combined into layouts pricing card, paywall

Dark mode The semantic layer remaps to different primitives. Components don't change.

By the time I left, the new product was generating and editing presentations through chat, the paywalls and the AI credit system were live, polls had shipped to a live A/B test, and the design system was in production across the marketing site and the new product surfaces. The design-system foundation made every UI effort after it faster and left the product more consistent and polished.