I treat design tokens as **expressions of intent**, not coding constants: primitives (OKLCH ramps, spacing units, motion curves) are never accessed directly—only semantic names like `surface.intent.primary` or `type.intent.emphasis` ship to iOS, Android, and web. Variable typography follows the same rule: weight, width, and optical size are abstracted into meaning-driven roles (`editorial.lead`, `utility.meta`) so light/dark mode changes don’t reinterpret style, only context. Platform outputs are generated from a single source of truth, with overrides reserved strictly for rendering physics, not design decisions. On one rollout, introducing an alias-only policy and renaming visually named tokens to intent-based semantics prevented a cascade of QA bugs when contrast ratios were adjusted late—dark mode stayed coherent across platforms without touching components. Because nothing referenced raw values, the system absorbed change without diverging on the different platforms. For art directors, thinking in terms of intent is the difference between *styling* an interface and *directing* it.