Token Sweep

What it is. A token sweep is Cursor's agent ripping through your entire codebase to replace every outdated design token with its new semantic equivalent in a single coordinated pass. You feed it the updated tokens file, point it at your Figma frame via MCP if needed, then give one clear instruction. The AI reads your full design system context, maps primitive values to semantic names, visits every component, updates variants, hover states, dark mode pairs, and even catches the hardcoded hex values engineers drop when they are shipping at 2 a.m. It runs your linter, formatter, and tests afterward. This is not autocomplete. This is maintenance at scale. In 2026 it became one of the four workflows that actually pay off because it kills the translation tax between what you ship in Figma and what users actually experience in production. The bigger and older your repo, the more valuable it gets. A 400 file Next.js app with three years of token drift is its natural habitat.

What it is not. A token sweep is not find and replace in VS Code. That method always misses the dynamically constructed class names, the template literals, and the one utility file no one remembers. It is not a Figma plugin or a visual update inside your design file. The sweep happens in the codebase where your real CSS-in-JS, Tailwind, Mantine, or Carbon implementation lives. It is not set and forget automation. The model will confidently break your layout if you do not review the diff. It is not for generating new UI from nothing. That is v0 territory. A token sweep operates on your existing, messy, production codebase and forces it back into compliance with your current rules. It is also not useful on brand new repos with five files and no established patterns for the AI to follow.

Concrete example. The brand team refreshed the primary blue in Q1 2026. The old primary-500 mapped to a flat #3b82f6 that failed the new contrast requirements. The designer opened Cursor, connected the Figma file through the MCP server to pull exact HSL values, and typed: Run a full token sweep replacing every deprecated primary token and matching hardcoded hex with the new semantic brand-primary-600 and brand-primary-hover. Update all button, link, chart, and navigation variants. Respect our Mantine component patterns and IBM Carbon dark mode implementation. Fix any overrides that no longer make sense. Format and lint everything.

The agent touched 187 files across the monorepo. In the Cal.com-style booking flow it updated the selected slot backgrounds, the confirmation panel accents, the sidebar indicators, and the email template that mirrored the web UI. It caught three components where engineers had ignored the token system during a launch crunch and brought them back into the fold. It left one promotional hero banner untouched because the sweep recognized the intentional brand deviation for that campaign. Total agent time was 21 minutes. The designer spent 14 minutes reviewing a clean diff before merging. The same task in 2024 would have taken three people and most of a sprint.

Another sweep hit a portfolio site built like Emil Kowalski or Rauno's work. The team shifted from arbitrary spacing values to a strict semantic scale. One prompt updated every gap, padding, and margin token while preserving the custom spring configurations in Motion.dev components. The AI correctly scaled the micro interaction timings so the hover lifts and transitions still felt expensive. A third example came during a typography audit on a Linear-style dashboard. The sweep migrated from fixed font classes to a fluid scale tied to new heading-1 through heading-4 semantic tokens. It updated not only the text but the related line heights, letter spacing, and even the badge components that inherited those values. Every Storybook example stayed in sync automatically.

When to use when not to. Use a token sweep after any design system change that needs to propagate without exception: rebrands, new radius scales, updated shadow palettes, or quarterly hygiene passes that keep your UI from rotting. They are the perfect first workflow for designers entering Cursor because the changes are mechanical enough for the AI to nail yet visible enough for you to judge instantly. Pair them with MCP so the values come straight from Figma instead of guesswork. The payoff is highest on mature codebases where token drift has become a tax on every new feature.

Skip token sweeps when the change requires nuanced human judgment on every instance. If updating the token alters visual hierarchy in ways that need product sign off, do the strategic decisions first then sweep the mechanical parts. Do not run them on tiny prototypes or right after a major framework migration when patterns are unstable. Never use them if you cannot read the JSX diff and test the live build. The AI remains capable of confident nonsense. A bad sweep creates subtle regressions that ship to production and erode trust.

Token sweep turns design system maintenance from an endless engineering ticket into your Tuesday afternoon superpower.