Danny Woo, Head of AI Products at Canva, introduces the session, sharing insights on Canva Code and the integration of AI, particularly with Anthropic's Claude models.
Canva’s mission is to empower the world to design, making the entire design process simple and accessible on one platform.
Historically, design was complex, costly, and fragmented across multiple tools; Canva unified these steps onto a single platform, supporting creation, publishing, printing, and more.
Canva is designed to support everyone, from students to enterprise creatives, across languages and devices.
AI supercharges Canva’s mission by enabling faster, easier, and more powerful design creation for all users.
Canva’s AI approach involves three pillars: developing proprietary models for unique design needs, integrating best-in-class commercial or open-source models, and offering an app ecosystem for developers.
The AI journey began in 2017 with machine learning for search ranking and color recommendations.
Key AI-powered features include background remover, Magic Write, Magic Edit, and Magic Design, integrated within Canva Magic Studio.
Recent launches include Canva AI, Canva Sheet, and Canva Code, expanding the AI feature set.
Canva Code uses Claude to generate interactive prototypes based on user prompts, enabling non-technical users to create websites and interactive content without coding knowledge.
The product idea emerged from the desire to transform the prototyping process, moving from confusion (“chaos”) to clarity when developing new product ideas.
Interactive prototypes were found to be highly effective at communicating ideas, but traditionally required significant engineering resources.
Internally, Canva teams leveraged Claude to rapidly create, share, and iterate interactive artifacts during product development.
Unlocking Prototyping and Empowering Non-Developers 07:46
Prototyping with Claude enabled more teams to make more prototypes, accelerating experimentation and feedback.
Embedding working prototypes in Canva designs demonstrated the tool’s potential beyond internal use, highlighting immense value for non-technical users (e.g., teachers, planners, marketers).
Only about 0.5% of internet users know how to code, so Canva Code aims to democratize interactive creation for the remaining 99.5%.
A functional prototype using Claude Sonnet 3.5 was quickly developed for real-world testing, allowing rapid iteration on UX and concept evaluation.
User testing with minimal guidance revealed natural use cases, confusions, and “magic moments,” feeding back into product improvements.
Canva Code supports embedding user-generated content (using iframes) in designs, mirroring existing approaches for supporting embeds in Canva.
Model Selection and User Experience Considerations 12:37
Model choice for Canva Code factored in not just correctness (evals), but also user delight, ability to infer intent from minimal prompts, web design quality, latency, scalability, and capacity.
Comparative testing showed Claude Sonnet models excelled at generating visually pleasing, interactive web content and SVGs, important for Canva’s user base.
The team highlighted the importance of evaluating models based on holistic user experience, not just benchmarks.
Clearly communicate AI limitations within the product, to prevent user confusion when generated code or features (e.g., contact forms) are only functional on the front-end.
Functional prototypes are invaluable: rapid prototyping outside the main codebase accelerates learning and innovation before formal development begins.
Differentiation is crucial: Canva Code focuses on nontechnical users, leveraging Canva’s integrated design editing to offer features (e.g., color pickers, brand fonts) that empower novices and unlock new levels of creativity.
Tips for generating beautiful UIs with AI involve careful prompt engineering: instruct models on design principles, color choices, and information hierarchy, embedding these in prompt defaults.
Future of AI in design is seen as a powerful enhancer of human creativity—AI reduces tedious tasks, enables rapid prototyping, and unlocks new possibilities for creative processes.
The session concludes with audience appreciation and closing remarks.