Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Coding tasks

Chrome's modern-web-guidance: skills for your agent

Written November 12, 2026

Coding agents keep writing yesterday's web code because their training data is full of it. Modern Web Guidance is a free, open set of skills that puts current browser practice, compatibility data and fallbacks into the agent's context at the moment it writes the code.

What it is

GoogleChrome/modern-web-guidance is a set of skills that embed web platform expertise, best practices and browser compatibility data directly into coding agents, steering them away from legacy patterns and toward native, high-performance platform solutions. The repository is stewarded by the Google Chrome, Microsoft Edge, Mozilla Firefox and MDN teams alongside the web development community. It is Apache-2.0 licensed. As read for this page, the repository shows about 6,980 stars and 239 forks, and it was created in March 2026. This is a standing resource, not a launch: the useful question is what is in it now.

What is covered

The guidance covers 250 modern web features, grouped in the README as CSS and Layout (140 features), HTML and DOM (36), and JavaScript and APIs (74). The core disciplines are user experience, CSS layout, performance, forms and UI, accessibility, and built-in AI, with guide categories running from passkeys and WebMCP to PWA, privacy, security and visual design. Each guide is written to be token-efficient: the project runs evaluations to prune material that models already know, so the agent's context carries the patterns it tends to get wrong, such as correctly structuring APIs it frequently misuses.

How an agent uses it

The quickstart is one command, npx modern-web-guidance@latest install, which runs an interactive wizard. Without installing, an agent or a person can try the CLI directly: search finds relevant guides, retrieve fetches one by ID, and legacy-scan scans a codebase for legacy patterns to modernize. In normal use the flow is three steps. The agent activates the skill for a relevant task. It then runs a local semantic search that matches the query to the best guide, using an offline, zero-dependency embedding model and BM25, with no network calls and no API keys. Finally, it retrieves the guide, so its code patterns, gotchas and fallbacks land in the context window.

Installation routes exist for most agent homes: the Vercel skills CLI (npx skills add), the GitHub CLI (gh skill install), Google Antigravity, the Copilot CLI, a Claude Code plugin, Grok Build, Kimi Code and Codex, each through that tool's own plugin or marketplace flow.

The part that makes it safe to adopt

Modern features are only useful if older browsers do not break. The guidance distinguishes purely additive enhancements, which older browsers can silently ignore, from critical behaviour such as dialog controls, where it writes considered, low-overhead fallbacks, generally under 50 lines or conditionally loaded polyfills instead of heavy bundles. It uses live compatibility data from the Baseline project so agents adapt to current browser support, documents quirks such as payload quotas, and gates guides for Origin Trial and other volatile features: they stay discoverable by search, but retrieval needs explicit developer consent, set through an optional .mwgrc configuration file that a team can commit and share. The project also publishes an evaluation harness that runs guided and unguided agents on the same tasks and compares pass rates and accessibility with and without the skill.

Limits

Related playbooks

Sources

Back to: Coding playbooks hub · Also: Addy Osmani's 25 agent skills for coding agents · Antigravity CLI 1.5.7: Auto becomes the default