Rosemary, made readable by agents
- client
- This site
- years
- 2026
- categories
- [systems] [ai]
- role
- Everything, from the tokens and components to the agent guidance and the eval
- access
- open
- markdown
- /ai/showcase/rosemary.md
- human
- /rosemary
## project
Rosemary is the design system this site is built on. I wrote it for the people and the AI agents who build with it, then measured whether the agents follow it.
## evidence
- guidance
- The project's agent instructions tell Claude which tokens, text styles, motion, layers and components to use, and what to do when something is missing.
- check
- A script flags any color, font size, spacing, radius, motion or z-index value that bypasses a token, and the site will not build until every one is fixed.
- eval
- I gave Claude 9 everyday requests, such as a filter, a dialog and a newsletter sign-up. Each one was built by a fresh Claude agent in its own copy of the site, once with the Rosemary guidance removed and once with it in place.
- result
- Off-system values fell from 6.3 to 0.0 per 100 CSS declarations. The agent used the right Rosemary component in 27 of 27 runs with the guidance and 9 of 27 without it. The site still built after 27 of 27 runs with the guidance and 26 of 27 without it.
- method
- Each task ran three times in each version with Claude Sonnet, and the same check scored both. I dropped a tenth task because the site already did what it asked. I built the eval with Claude Code.