Ask a coding agent for a modal and there's a good chance you'll get a div, a backdrop div, a focus-trap library and forty lines of JavaScript, when <dialog> does most of that for free. Modern Web Guidance exists to fix that habit. It's a skill pack and CLI that feeds your agent short, curated guides on modern web features right when it's about to write UI code, and version 1.0.0 was published to npm on 10 November 2026. Here's what's in it, how to install it for your agent, what a guide actually looks like, and where I'd be careful.
What Modern Web Guidance is
The project README describes it as "a set of skills that embed web platform expertise, best practices, and browser compatibility data directly into your coding agents," stewarded by the Chrome, Edge, Firefox and MDN teams. The problem it targets is stated plainly:
Coding agents often default to older patterns because LLM training data contains vast amounts of legacy code.
The current catalogue covers 250 web platform features and 203 real-world use cases, grouped into areas like accessibility, CSS, forms, performance, security, UI components and built-in AI. The guides are written to be token-efficient: the team runs evals and prunes material models already know, so your agent's context isn't flooded with basics.
The 1.0.0 tag landed on npm on 10 November after a long run of 0.0.x releases, so this is the first version that calls itself stable.
Installing it
The package needs Node.js 22 or newer, per its engines field on npm. Chrome's documentation lists several routes; the recommended one is an interactive wizard:
npx modern-web-guidance@latest install
Or install it for a specific agent. For any agent that supports the skills format:
npx skills add GoogleChrome/modern-web-guidance
For Claude Code:
claude plugin install modern-web-guidance@claude-plugins-official
For GitHub Copilot CLI, inside a session:
/plugin marketplace add GoogleChrome/modern-web-guidance
/plugin install modern-web-guidance@googlechrome
Try it before installing anything
You can run the CLI directly to see what an agent would see:
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 .
search returns guide IDs with descriptions, retrieve prints a guide, and legacy-scan looks through a codebase for patterns worth modernising. Run the search on a couple of components you built recently. It's the quickest way to judge whether the guides match how you'd want them built.
How your agent uses it
The skill's instructions (its SKILL.md) tell the agent to run it first for HTML, CSS and client-side JavaScript work, and explicitly not for backend, infrastructure or tooling tasks like Vite configs or tests. The workflow is:
Search with a short, generic query of two to five words, such as "light dismiss dialog", leaving out framework names. "Close React settings modal on outside click" is given as the wrong way.
Retrieve the matching guide or guides.
Verify the code against the guide before finishing.
Disclose: the agent ends its answer with a "Modern Web Guidance" section saying which guides shaped the code and why.
That last step is the one I like most. You get a short note explaining, say, why it used showModal() instead of a custom overlay, which turns the agent's choices into something you can review.
What a guide looks like
The navigation-drawer guide, linked from the README, is a good example of the style. It builds the drawer as a modal <dialog> opened with showModal(), so the browser makes the rest of the page inert, handles Escape and restores focus. The swipe-to-close gesture uses CSS scroll snapping instead of pointer-event JavaScript, and the backdrop fade is tied to scroll position with a scroll-driven animation. The markup starts like this:
<dialog class="Drawer" id="drawer" aria-label="Navigation drawer">
<div class="Drawer-scroller">
<nav class="Drawer-sheet">
<button id="drawer-close" aria-label="Close menu" autofocus></button>
<!-- links -->
</nav>
</div>
</dialog>
The guide marks the accessible name as mandatory and explains when to use a popover="manual" element instead (when the same panel also acts as a sidebar on wide screens). That's the level of judgement you want an agent to inherit. If you've followed our practical guide to React 19.3 view transitions, you'll recognise the general direction: let the platform do the animation work.
Where to be careful
Check the output against your own browser targets
Jon Henshaw, writing at Coywolf, warns that "letting it control how everything is coded can have unintended consequences." In his case it added scripts for legacy browsers and code loaded from third-party servers. His approach is to keep a reviewed local copy as a skill, updated weekly with his own rules. Whether or not you go that far, decide your browser support policy first; our note on which new CSS features reached Baseline is a reasonable place to start that conversation with your team.
legacy-scan is partial
The skill's own instructions say legacy-scan checks a curated subset, roughly 20% of the guidance corpus, and that findings must be checked for false positives. A clean scan doesn't mean a modern codebase.
Keep agent permissions narrow
The tool fetches from npm, so your agent needs network access to run it. The SKILL.md advice is sensible: allowlist npx -y modern-web-guidance@latest * specifically, "never bare npx *." Giving an agent blanket permission to run any npm package is a much bigger decision than installing one skill.
Should you install it?
If you use an agent for front-end work, yes, at least to try. The guides are opinionated in the right direction (native elements, less JavaScript, accessibility built in), and the disclosure step makes the agent's reasoning reviewable. If your project supports older browsers or has strict rules about third-party code, treat it the way Henshaw does: a reviewed reference rather than an autopilot. The same principle behind writing design specs that both people and agents follow applies here: agents do better with good rules, but someone still has to own the rules.



