Chrome's Modern Web Guidance Gives Coding Agents 250 Web Features to Use Instead of Libraries

Google's Chrome team has published Modern Web Guidance, an Apache-2.0 repository of agent skills that tells coding agents to reach for native browser features before a library, and it covers 250 web features and 203 use cases as of Nov. 11.

The repository, GoogleChrome/modern-web-guidance, says it is "stewarded by the Google Chrome, Microsoft Edge, Mozilla Firefox, and MDN teams, alongside the web development community." Its reason for existing is a training-data problem: "Coding agents often default to older patterns because LLM training data contains vast amounts of legacy code."

What is in the 250-feature pack

The README splits the 250 features into CSS and layout (140), HTML and DOM (36), and JavaScript and APIs (74). The 203 use cases are grouped under labels such as dialog, passkeys, navigation-drawer and scroll-snap-state-sync.

Feature group	Features
CSS and layout	140
JavaScript and APIs	74
HTML and DOM	36

The README says the guides are "designed to be token-efficient" and that evals are used to cut content models already know. Several entries carry a tmp- prefix, for example tmp-webcrypto-modern-algos, which suggests some guides are placeholders.

1. CSS and layout140 features
2. JavaScript and APIs74 features
3. HTML and DOM36 features

Source: GoogleChrome/modern-web-guidance README, accessed 2026-11-11

How to install it on Claude Code and four other agents

Chrome's documentation page gives five routes. The recommended one is npx modern-web-guidance@latest install. Claude Code users run claude plugin install modern-web-guidance@claude-plugins-official, and Copilot CLI users add a marketplace named googlechrome. Antigravity CLI and Vercel's npx skills add GoogleChrome/modern-web-guidance are the other two.

The README also lists search, retrieve and legacy-scan commands. The documentation page names only install, so the scanner is documented in one place and not the other.

The sample prompts on the Chrome page are concrete: move modals to the <dialog> element, migrate tooltips to the Popover API, add passkey login, and cut long tasks that hurt Interaction to Next Paint.

What one demo shows about a single prompt

The only worked example found so far is a LogRocket tutorial by Emmanuel John, dated July 22, which ran Gemini CLI against two projects with the same prompt. Without the skills, the agent wrote a modal with a custom portal and a 300 ms timer that had to match the CSS transition by hand. With them, it used a native <dialog> element.

The guided agent also switched task cards to CSS container queries and a tooltip from Floating UI to CSS Anchor Positioning. For a search over 2,000 tasks it used scheduler.yield() in chunks of 50.

John is explicit about the limit: it is "a single demo, not a controlled benchmark." He also wrote that native APIs do not always beat libraries, since Anchor Positioning needs a fallback for older browsers and native dialogs still need manual keyboard and screen-reader testing. He said Chrome's documentation describes the project as an early preview. The Chrome page itself gives no version number or release date.

What the Chrome page does not say

Neither the README nor the Chrome page publishes an accuracy figure for the guides, a list of tested agent versions, or a count of how often an agent follows them. The default target is Baseline Widely available, so a project with older browser requirements has to set its own target.

Source code and issue tracking live in a separate repository, GoogleChrome/modern-web-guidance-src, so bugs filed on the main repository go nowhere. Chrome's page points bug reports to Google's issue tracker under component 1400036.

The pack joins a crowded shelf of repositories that package instructions for agents. The Terminal compared two of the largest earlier this month, and the same agent-tooling wave produced Vercel Labs' agent-browser. This one differs in its stewards: it carries browser vendors' names on the README.

The next thing to watch is the tmp- entries, which the README does not say will be finished or removed.