I lost three months arguing with prompts before I gave up. Every screen Cursor wrote looked roughly right in isolation. Put any two side by side and the cracks showed: a 14px gap here, 16px there, a button radius that drifted from 6 to 8 to 10 across the codebase. By the tenth page nothing matched.
The fix isn't a smarter prompt. The fix is a system the AI can read from.
Why prompts decay
A prompt is local. It tells the model what to do for the next file. The model has no memory of the rest of the project. Even with project-wide rules in your AGENT.md or .cursorrules, those rules are prose — 'use a calm tone', 'prefer rounded corners' — open to interpretation every single generation.
Design tokens are the opposite of prose. They are values. `--radius-md: 8px` is not interpretable. Either the model uses it or it does not, and a static analyzer can tell you which one.
Three primitives, one language
Windcraft splits the design language into three reading levels for the AI:
- Tokens — atoms. Color, spacing, type, radius, shadow, motion. Edited in the dashboard, synced to Tailwind config + CSS vars + RN theme.
- Contracts — grammar. 77 component recipes (Button, Card, Modal …). Each defines variants, sizes, slots, rules, examples.
- Patterns — sentences. 60 page recipes (sign-in, checkout, dashboard …). Layout regions, states, microcopy, motion, haptics.
All three feed into Cursor, Claude Code, Codex, Copilot via MCP. The AI doesn't have to remember anything: it asks the server, gets a structured answer, and writes against it.
The dogfooding test
We rebuilt our own dashboard in a week using the V1 product. The diff: same six tokens cascading through 30 files, no inline rgba(0,0,0,0.06), no rounded-md vs rounded-[8px] vs rounded-2 anywhere. Every primary button is the brand colour. Every primary button is the same brand colour.
If you can't generate the system, you can't enforce it. If you can't enforce it, AI will erode it one prompt at a time.
Prompts are how you ask. Tokens are how you answer.