With Modern Web Guidance, Claude Code running Opus 5.5 goes from passing 57% of assertions to 88% across the 215 web tasks Google uses to evaluate it. The skill, run by the Chrome team together with Edge, Firefox and MDN, shipped version 1.0.0 of its CLI to npm on November 10, 2026, and hit the Hacker News front page the same day with a 278-point Show HN. It has 3,342 GitHub stars and 40,300 installs on skills.sh.
What it does
The problem is familiar: models are trained on years of old JavaScript and CSS, so when you ask for a modal or a responsive table they reach for 2019-era libraries and hacks instead of <dialog>, the Popover API, anchor positioning or container queries. The skill feeds the agent expert-written guides on 250 web platform features (140 for CSS and layout, 36 for HTML and DOM, 74 for JavaScript and APIs), with browser support data from Baseline and the recommended fallback when a feature isn't everywhere yet.
Its SKILL.md declares itself mandatory for any HTML, CSS or client-side JavaScript task, and deliberately excludes backend, infrastructure and tooling work. The flow has four steps:
- Search: the agent runs
searchwith a 2–5 word query about a single capability ("light dismiss dialog", not "close React settings modal"). - Retrieve the guide with
retrieve, plus the general guide for the discipline (accessibility, performance, forms…). - Verify the generated code against the guide.
- Disclose: end the response with a
### Modern Web Guidancesection listing each guide applied and the specific rule that prevented a mistake.
Search runs locally: an embedding model and BM25 ship inside the package, with no API calls and no keys. There's also a legacy-scan mode for auditing existing code, which only covers about 20% of the catalog (the guides with deterministic rules). Origin Trial APIs and "volatile" features (single-engine, shipped within the last three months) show up in search, but the agent won't retrieve them unless you opt in through a project .mwgrc file.
Installing and trying it
You can see what it returns without installing anything:
npx modern-web-guidance@latest search "animate a dialog modal backdrop"
npx modern-web-guidance@latest retrieve "animate-to-from-top-layer"
npx modern-web-guidance@latest legacy-scan .To install it in Claude Code as a plugin:
/plugin marketplace add GoogleChrome/modern-web-guidance
/plugin install modern-web-guidance@googlechromeIt also installs with npx skills add GoogleChrome/modern-web-guidance, gh skill install, or the plugin commands for Codex, Copilot CLI and Antigravity listed in the README.
What it needs and what it risks
- It runs code from npm on every query. Each search is an
npx -y modern-web-guidance@latest: your agent downloads and runs the latest published version without asking. The package has no dependencies (which shrinks the supply-chain surface) and is about 16.5 MB unpacked, but@latestmeans trusting every future release. The skill itself asks you to allowlist onlynpx -y modern-web-guidance@latest *, never a barenpx *. - It needs network access the first time to fetch the package; searches themselves are local.
- It doesn't touch credentials or send your code anywhere.
- It's pushy. The skill description says "MANDATORY: Execute FIRST" for any front-end task. On a UI-heavy project that means several calls and guides in context per task, plus an extra block at the end of every response.
- It doesn't guarantee accessibility. Accessibility specialist Adrian Roselli tried the first example in the docs (an animated accordion) and got code whose animation didn't work in Firefox and that failed WCAG. His core point: Baseline tracks what browsers support, not what assistive technologies support. That review dates from May 2026 and the project has moved on since, but the argument still holds.
Compared with what we had
| Modern Web Guidance | Vetted pattern library | No guidance | |
|---|---|---|---|
| How it reaches the agent | Skill + local CLI | Docs or components you copy | Whatever the model remembers |
| Deterministic | No: the agent picks the guide | Yes | No |
| Keeps up with the platform | Updates with each package release | Depends on the maintainer | Model's training cutoff |
| Evidence of improvement | Public evals: +28 to +36 points | — | Baseline: 52–60% |
The README evals are its strongest argument: on November 8, 2026, Antigravity with Gemini 4 Argon went from 58% to 94%, Codex with GPT-6.1 Sol from 54% to 82%, and Claude Code with Opus 5.5 from 57% to 88%. These are the team's own tests, using Playwright graders on sample apps, but the dashboard is public.
My take: if your agent writes front-end code, install it, because the gap versus nothing is large. What it won't do is replace a vetted, accessible component library: for a production modal, menu or accordion, start from something tested and let the skill keep the rest of your code out of 2019.
Sources
- GoogleChrome/modern-web-guidance (README and SKILL.md), GitHub.
- Show HN: Get your coding agent off ~2019 JavaScript, onto modern web primitives, Hacker News, November 10, 2026.
- modern-web-guidance on skills.sh.
- modern-web-guidance on npm.
- Maybe Don't Rely on Google's "Modern Web Guidance", Adrian Roselli, May 25, 2026.