AI Design System Infrastructure
Role
Design Engineer
Duration
Continuous project
Responsibilities
MCP Architecture, agent engineering, DS structure
Focus
AI-First & Infrastructure
CONTEXT
The need to rethink interface construction
During the internal restructuring focusing on broad AI usage, the need arose to transform design into an assisted system. Screen creation still relied on manual processes, with high effort in operational UI decisions and frequent inconsistencies.
Manual Decisions
High effort in repetitive UI details.
Inconsistency
Design system variations occurred frequently.
THE PROBLEM
The process had clear limitations: slow and repetitive creation, reliance on tacit knowledge, and frequent rework due to misalignment. This directly impacted delivery speed and the team's strategic focus.
OBJECTIVE AND APPROACH
The goal was to build an infrastructure that automated Figma layouts via natural language. We decided on an MCP server-based system, where specialized agents perform specific tasks and Figma is controlled programmatically.
ARCHITECTURE DECISION
The system's core is an MCP server responsible for serving structured context to agents, centralizing rules and ensuring consistency between design and development.
MCP Architecture
Single source of truth for design rules.
Vibma + Figma
Direct programmatic control over the interface.
TOKEN EFFICIENCY (CRITICAL DECISION)
Token consumption was treated as a core constraint. We implemented three response modes to optimize performance and reduce latency: ids, summary, and full.
AGENT SYSTEM
The workflow was divided into specialized agents (Intake, Build, Review, Learn) based on the actual design process, avoiding context overload and increasing precision.
AI IN PRACTICE
The entire system was developed using Claude as active infrastructure: from architecture definition to prompt engineering and catalog structuring.
RESULTS
Speed
Creation reduced from 1 day to ~2 hours.
Quality
Increased consistency and focus on product decisions.
Scalability
Architecture also accessible to developers.
MY ROLE AND LEARNINGS
I was the architect and executor from end to end. The main lesson was that design systems gain real scale when treated as technical infrastructure, where token efficiency dictates system viability.