Short thoughts from the edge of product, design, and AI. Observations made while building at the frontier, not from a safe distance.
Tokens were built for developer handoff pipelines. LLMs reason differently. When an AI encounters a token like color.brand.primary, it has no semantic grip on what that means. Here's how I've been rethinking the contract.
Most teams are asking "how do we ship AI features?" They should be asking "how do we structure our decisions so an AI can reason about them?" Readiness is an architecture problem, not a tooling problem.
Every design system makes implicit promises about consistency, about trust, about who holds the decisions. Most systems break down not because of technical debt, but because that contract was never made explicit in the first place.
When the output is unpredictable, the input architecture matters more. Generative tools don't eliminate the need for systems — they give systems a new structural role. The constraints are where the craft lives.
When AI can generate hundreds of design variants in seconds, the designer's job becomes defining which constraints make an output worth keeping. Taste is still the variable — it's just expressed upstream now.
When you ask an LLM to build a UI component and it goes wrong, the failure usually isn't the model. It's the absence of clear, structured intent in how the component was documented to begin with.