The shadcn-ui organization released @shadcn/lint, an open-source linter for Tailwind-based design systems built to give AI coding agents rules they can check their own work against, according to its GitHub repository.
Teams define what styling changes are allowed on a component without editing the component itself. When an agent breaks a rule, the linter returns an explanation and a fix based on the project's own components, variants and theme values, rather than a generic type error, the repository says. It supports React, Vue and Svelte, and runs through ESLint or Oxlint.
In testing described in the repository, agents using models including Claude Sonnet 5 and Opus 5 completed 8 of 8 assigned tasks, and lint violations dropped from a range of 42 to 117 errors down to zero once the agent had the linter's feedback to work from. Fixing violations cost 10% to 48% less with that feedback than without it, according to the repository.
Most linters assume a human reads the error and knows what to do with it. One built to hand an agent a fix tied to the project's actual design system, instead of a generic rule, is a small but concrete example of tooling built agent-first rather than retrofitted for agents.