Semantic tokens
Middle layer of a 3-tier token system: intent + context. Examples: `color-action-primary`, `color-feedback-error`, `spacing-component-button-padding`. Per Nathan Curtis: *'semantic tokens as the choices'*, the subset of primitives actually used, with a name that says WHY. Romina Kavcic pushes in 2026 for explicit DTCG `$description` per token.
Strengths
- Layer components reference, single leverage point for rebrand
- With DTCG `$description`, becomes LLM-readable, the agent knows when to use this token
- Compatible with DTCG mode×brand×theme resolvers, semantic stable, primitive switches
Limitations
- Naming is teams' #1 pain point, `global`, `semantic`, `primitive`, `alias` overlap and lose people
- Risk of becoming a flat alias without intent, Nate Baldwin anti-pattern
Best for
- Any serious DS, the semantic layer is not optional
- DS exposed to AI agents, explicit intent drastically reduces fabrication