- GitHub says it moved GitHub.com styling from CSS-in-JS to CSS Modules, reporting 55% less server render time and 25% less component start-up time on one unnamed page after the Primer phase.
- These are GitHub's own figures for one page, and a later phase reported only 1% to 22% server-render gains on some pages, so the two are not comparable.
- This is one company's account, not a benchmark; check your own load metrics, test coverage, feature-flag tooling and hidden dependencies such as theming before promising a migration date.
GitHub says it rebuilt its styling system without breaking the site
GitHub's Primer design system team reports that it moved component styling on GitHub.com from CSS-in-JS to CSS Modules. In this format, styles are bundled into ordinary stylesheets delivered with the page's HTML, so no styling logic has to run in the browser or on the server. The account, published on the GitHub engineering blog on September 27, 2026, says the trigger was a 2023 jump in the number of components on certain pages. GitHub names three side effects of its earlier setup: slower first loads, because styling had to be set up in the browser; weaker server rendering; and style changes that became hard to manage. The post does not say when the migration itself began.
GitHub states that it completed the work "without breaking GitHub along the way." That is the company's own characterisation. The same post mentions "a few hiccups" during the later theming phase and credits feature flags and production stress-testing with helping the team catch and repair errors.
The 55% is a page-level figure from the Primer phase, not the whole programme
Both figures are GitHub-reported results for the Primer component migration, finished in December 2024, and both are stated for a page rather than for GitHub as a whole. Each component moved behind a feature flag, was checked against existing visual regression tests, and was rolled out first to the team, then to GitHub staff, then to all users. The later phase, removing the sx customization prop from GitHub's own codebase, has a separate, lower reported range: server-side rendering improvements of 1% to 22% on some pages. The post does not say whether the same pages were measured in each phase, so the two sets of figures are not directly comparable.
That second phase began in April 2025 with a peak of about 7,760 sx props. GitHub describes it as "mostly automated," using a VS Code plugin and a codemod, with some manual oversight and validation. A rotation of eight engineers migrated 6,419 props over six months. The post does not say they worked on it full time.
Two engineers and Copilot agents cleared the last props; theming took two more months
By April 2026, 895 sx props remained. GitHub reports taking that number to zero in three weeks with two engineers and "a whole lot of Copilot coding agents." The post does not isolate how much of that pace came from Copilot. The remaining props may have differed in difficulty from the first 6,419, and the plugin and codemod had already been built, so comparing the two phases by headcount and elapsed time is not like for like.
Clearing sx did not end the project. GitHub supports seven themes, each with a high-contrast variant, all enabled through styled-components, so theming had to be decoupled first. That took two more months, again with feature flags and a few hiccups, and the dependency removal was itself put behind a flag. By GitHub's account, it has run entirely on CSS Modules since June 2026, and sx, styled-components and styled-system are gone from its dotcom code.
What a budget-holder can take from one company's account
This is a single company's report, not a benchmark for cost or duration elsewhere. GitHub's in-house tooling and engineering scale are specific to it. What the post does support is a set of dated milestones: Primer components done by December 2024, the sx phase from April 2025 to about May 2026, and 100% CSS Modules in June 2026. It also shows where GitHub placed its risk controls: feature flags, visual regression tests and staged rollouts at each step. Those belong in a migration budget from the start.
Questions for the engineering team: Do our server-rendering and page-load metrics show whether the current front-end styling approach is costing load time, as GitHub reported for CSS-in-JS? If AI coding agents are proposed for a mechanical migration, what test coverage and feature-flag tooling exists to check their output, and how would we measure their contribution separately from existing codemods? And which dependencies, such as theming, sit behind the obvious target and have been scoped in the plan before a completion date is promised?
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: github.blog.





