Let’s be honest: most AI for design has been about generating a pretty static image and hoping you can recreate it. But for professional designers, the real work isn't in the 'big reveal'—it's in the tedious, iterative grind. We're talking about adjusting spacing by 4px, tweaking a hex code, or reworking a layout based on a client's vague feedback. That's where Figma is stepping in with a game-changing approach: a model that actually speaks the language of the canvas.
Moving Past the Generalist
For a long time, frontier AI models struggled with Figma's complex syntax. Trying to force a general-purpose LLM to edit a design often required massive system prompts that burned through tokens and slowed down the workflow. To fix this, Figma shifted from a 'generalist' approach to a 'specialist' one.
By focusing on specialized post-training, Figma built a model specifically fluent in design editing. Instead of just guessing what a UI should look like, this model understands the actual structure of the Figma canvas, allowing it to perform precise, non-verifiable edits without interrupting the designer's flow.
The Secret Sauce: 250,000 Examples
How do you teach a machine to 'feel' a layout? Figma leaned into data. They trained their early specialist model on over 250,000 examples, combining starting designs, synthetic requests, and successful edits in Figma's native format.
The team evolved their strategy from supervised fine-tuning—which teaches the model by example—toward reinforcement learning. The result is a first-party model integrated into the Figma agent that can compete with the world's leading frontier models on specific canvas tasks, providing a snappier, more intuitive editing experience.
The Future of the Canvas
This isn't about replacing the designer; it's about removing the friction. By dividing the work between specialized first-party models and third-party generalists, Figma is turning the AI agent into a true collaborator. We're moving toward a world where you don't just prompt a design into existence—you iterate with an AI that knows exactly where the padding is wrong.
Sources
Media
