AI Agents' Design Consistency Problem

AI coding agents generate clean, modern layouts with sensible colors on individual pages, but fail to maintain shared logic across an app. Result: mismatched buttons, drifting grays, inconsistent spacing—polished singly, incoherent together. Author's real-world example: after five pages in a new app, each had its own implicit design system. CLAUDE.md files guide agents on coding behaviors (like the 4 lines bookmarked by 60,000 developers, addressing issues Karpathy diagnosed), but ignore visual design rules, so agents never inquire or apply them.

Trade-off: Agents prioritize functionality over aesthetics without explicit constraints, leading to 'fine on its own' pages that don't scale.

Google's DESIGN.md: Markdown Spec for AI Design Guidance

Open-sourced by Google on April 21 via a blog post on Stitch DESIGN.md, this plain-text markdown format describes full design systems in a way AI agents parse natively—no APIs, dependencies, or build steps needed. It directly solves the wall every designer/developer hits: code works, pages don't match.

Core claim: Every DESIGN.md requires exactly 9 sections to constrain agent behavior, preventing UI drift and enforcing consistency. (Specific sections detailed in full article; spec enables agents to 'read' tokens, components, layouts like humans.) Impact: Turns autonomous UI generation from chaotic to production-ready, letting small teams ship visually coherent apps faster.