The Evolution of Design Systems in the LLM Era
Moving beyond static component libraries toward dynamic, AI-governed design architectures.
1The Consolidation Challenge
In 2022, the author undertook a project to consolidate fragmented design kits into a single company-wide system. The goal was to unify disparate design languages across various business units. Despite the effort, the resulting system became massive, containing 1,786 variants, illustrating the inherent difficulty of managing complex design requirements manually.
2Figma Limitations and Bloat
The system relied heavily on Figma variants to handle different sizes and states, leading to a bloated file structure. At its peak, the working file contained 4,000 variants. This approach was a technical necessity at the time, as native variable theming and advanced component slots were not yet fully available.
3The Arithmetic of Components
A single button component required 260 variants to cover three behaviors and five sizes. Only 15 of these variants required genuine design decisions regarding spacing and proportions; the remaining 245 were purely arithmetic. This highlights how manual component creation often confuses simple mathematical scaling with actual, meaningful design decision-making processes.
4Technological Shifts in Design
Figma’s introduction of typography variables in April 2024 would have drastically reduced the system's complexity. By utilizing variables instead of manual variants, the total count could have been cut by up to 80 percent. This shift demonstrates how modern tooling can replace labor-intensive manual asset management with more efficient, token-based logic.
5Governance via Design Councils
The project utilized a design council with representatives from each business unit to oversee decisions. This decentralized approach required significant communication skills and an altruistic mindset to succeed. Effective governance in such a structure demands time and a shared culture, as democratic decision-making is inherently slower than centralized, top-down design management.
6The Role of LLMs
The author proposes that governing an LLM skill could serve as a superior foundation for evolving design systems compared to static documentation. By treating design limits as a skill rather than a rigid kit, teams might better manage complexity. This approach could potentially automate the creation of components and layouts on-the-fly.
7Code as the Truth
Historically, the author viewed code as the only true source of truth for design systems. However, with the rise of AI, the definition of 'truth' is shifting. If an LLM can generate assets for any software via MCP, the traditional reliance on a single, static codebase may become less relevant over time.
8Lean Design Philosophy
Modern design systems benefit from a lean, progressive perspective that avoids baking every requirement into the core components. By using crude, simple base components that remain flexible, designers can accommodate variations without sacrificing functionality. This shift away from 'all-encompassing' kits allows for greater adaptability as product requirements evolve over time.
9Lessons from Retrospection
Looking back, the author notes that the consolidation work was a valuable learning experience despite its challenges. The project shaped their current approach, emphasizing that design systems should be enablers of problem-solving rather than rigid constraints. The experience highlights the necessity of balancing technical limitations with long-term, scalable design strategy.
10Future of System Design
The future of design systems lies in moving beyond static assets toward dynamic, AI-assisted architectures. As tools like Figma continue to evolve, the focus must shift from crafting thousands of variants to defining clear, governed boundaries. The next frontier involves proving how LLMs can effectively manage these systems at scale.
Original article: UX Collective - Medium — The Evolution of Design Systems in the LLM Era