What problem does it solve? It prevents AI-generated interfaces from defaulting to a generic average of training data by forcing one human aesthetic decision — what the product looks like and what it deliberately is not — and recording it where restricted-context subagents can inherit it. ## Core Features & Use Cases - Human-gated direction decision: Generates exactly three named design directions (type pairing, seed hue, shape language, density, signature element) and requires the user to pick one, recording the choice and the two discarded alternatives as an ADR. - Design system authoring with provenance: Writes design_system.md with three token layers, a full component roster, and contrast ratios computed by design-lint.js — estimated numbers are forbidden. - Three channel levels: Pulls components on demand from Claude Design (level A), draws on a native canvas with save-capability probing (level B), or runs fully offline with throwaway HTML comps (level C). - Use Case: A project with a configured frontend but no docs/03-ux-ui/ deliverables invokes this phase to produce a navigation map citing real contract operationIds and a verified design system before any page is built. ## Quick Start Ask the agent to start the UX design phase for the project so it generates the brief, three visual directions, and the design system documents.
