The author argues that Figma's complex design system infrastructure is becoming obsolete as the source of truth shifts back to code. Figma's proprietary primitives - components, styles, variables, and props - were designed to systematize design within engineering organizations, but their undocumented format excluded them from AI training data. Meanwhile, Claude Design embraces HTML and JavaScript as its foundation, giving it a structural advantage.

Figma's victory over Sketch came from establishing its tooling as canonical, but this created a locked-down ecosystem that's painful to work with programmatically. The author demonstrates this complexity with examples from Figma's own design system files, showing 946 color variables with eight mode-specific values each and components with 16 variants. Debugging such systems involves tracing through nested components, library swaps, and mode overrides.

As LLMs were trained on code rather than Figma's proprietary format, the author predicts design tooling will fork into two paths: tools like Claude Design that work directly with code, and pure exploration environments unconstrained by systems. The friction between design and implementation that has existed since the beginning of time is disappearing as tools like Claude Design and Claude Code become integrated, allowing direct conversation between design and implementation.