Ask your AI coding agent to add a scroll animation and it will reach for GSAP. Ask for a lazy-loaded image and it installs a third-party library. Ask for a tooltip and you get a 200-line JavaScript class — when the browser’s Popover API does it in four lines. This is not the agent being careless. It is being faithful to its training data, which reflects a web from the early 2020s — a web of polyfills, workarounds, and npm packages for capabilities that browsers have handled natively for years. Google shipped Modern Web Guidance 1.0 on November 10 to address that directly.
The Training Data Problem
The core issue is structural. AI coding agents learn web development from the full history of the web, and the early-2020s web is well-represented in that training data. That era’s code is full of polyfills for Array.flat, fetch, and structuredClone — features that every modern browser has supported for years. Research found that a model trained through June 2026 still generated old-API code for eight API changes that shipped before its own cutoff date. The agents are not broken. The signal they were trained on is simply old.
The practical result is predictable: agents default to GSAP for scroll animations instead of CSS scroll-driven animations, ship IntersectionObserver polyfills for a browser API that has been available since 2016, and write JavaScript tooltips when the HTML Popover API eliminates the need entirely.
What Modern Web Guidance Does
Modern Web Guidance installs as an agent skill — 203 use-case guides covering roughly 250 web platform features across CSS and layout, JavaScript APIs, HTML and DOM, accessibility, performance, and built-in browser AI. The skill integrates directly into the agent’s context, steering code generation toward native browser capabilities during the conversation rather than requiring the agent to look up a URL mid-session.
After installation the package works fully offline. The guides download locally and are read from the filesystem — no live calls to Google during coding sessions.
The governance model is cross-browser. Version 1.0 formalizes six owner seats: two from Chrome, one from Microsoft Edge, one from Mozilla, and two independent web platform experts. Core guidance must remain vendor-agnostic; browser-specific details ship only in separate opt-in packs. The guidance is built on web standards, not Chrome preferences.
The Numbers
The team runs nightly evaluation harnesses against real coding tasks. In runs completed November 8 — between 199 and 215 tasks per agent — guided agents passed 88% of automated checks. Unguided agents passed 56%. That is a 32-percentage-point gap on the kinds of tasks that appear in every frontend project.
In a production deployment at note inc., Modern Web Guidance raised pass rates from 61% to 94% across a 180-run evaluation on the company’s actual codebase. LCP image prioritization improved by 28% across 431 merged pull requests. Agents replaced custom JavaScript scroll animations with native CSS and migrated tooltips and accordions from JavaScript to HTML.
New in 1.0: The Code Audit
The most useful addition for teams with existing codebases is the legacy-scan command. It walks your HTML, CSS, JavaScript, and package.json looking for places where older patterns can be safely replaced with native browser features — without changing behavior.
npx modern-web-guidance@latest legacy-scan .Google acknowledges this is a first iteration using static analysis over a limited pattern set. It will not find everything and will not reason deeply about code yet — honest framing for a genuinely useful starting point.
Installing It
The interactive installer handles setup for whichever agent you use:
npx modern-web-guidance@latest installSupported environments include Claude Code, GitHub Copilot, Gemini CLI, Cursor, and Devin. No paid tiers, no authentication required. The source is on GitHub under the GoogleChrome organization, and the npm package installs without authentication.
What to Watch
203 guides is the starting point, not the ceiling. The nightly eval harness means guidance updates as browser features ship and mature. The governance model is new territory — this is effectively the first living, cross-browser agent skill standard for the web platform, with Mozilla and Edge participation keeping it honest on compatibility.
Future versions are expected to extend coverage to built-in browser AI APIs: the Prompt API, Summarizer API, and similar capabilities landing in browsers now. If this model proves out, it could become the standard mechanism for keeping coding agents current with any fast-moving platform, not just the web.
Install it, run the legacy scan, and find out how much of your codebase your agent has been keeping in 2019.
