Chrome's Modern Web Guidance Reaches 9,200 Stars While Its Docs and README Count 100 and 203 Use Cases
A.I. / news
Chrome's Modern Web Guidance Reaches 9,200 Stars While Its Docs and README Count 100 and 203 Use Cases
The agent skill set from the Chrome team reports a 37-point average gain in best-practice adherence, but publishes the method only on a dashboard.
Chrome's Modern Web Guidance, a set of skills that feeds web platform advice to coding agents, had 9,200 stars on GitHub on November 13, and its source repository gained 1,875 stars in a single day. The project's own documents disagree on how much it contains, and the only measured result is an early-results figure with no published method in the text.
The documentation says the guidance covers "over 100 use cases". The GoogleChrome/modern-web-guidance README lists 250 web features and 203 real-world use cases.
What it does
An agent searches by task with npx modern-web-guidance@latest search "<your task>", which returns JSON with each guide's id, description, token count and a similarity score. It then calls retrieve with an id to pull the Markdown guide into its context. The aim is to stop agents from writing legacy patterns, usually JavaScript-heavy ones, where modern CSS or HTML would do.
The source repository, which holds the authoring and evaluation code, had 4,100 stars, 240 forks and 307 open issues. Its README says the project is stewarded by the Chrome, Microsoft Edge, Mozilla Firefox and MDN teams, and that evaluations are used to prune content "that models already know" so guides stay token-efficient.
- CSS and Layout140 features
- JavaScript and APIs74 features
- HTML and DOM36 features
Source: GoogleChrome/modern-web-guidance-src README, accessed 2026-11-13
The one number is an average with caveats
The docs report that early results show a 37 percentage point average improvement in adherence to modern best practices when agents use the guidance. They add that results vary by project requirements, model, prompts and tools, and that evaluations use Playwright scripts and run several times a week. The text gives no model list, task count or baseline score.
LogRocket's July 22 write-up by Emmanuel John cites no benchmark. It compares two Next.js task-manager apps built from one prompt with Gemini CLI, one with the guidance and one without. The unguided build used @floating-ui/react for a tooltip where the guided one used CSS Anchor Positioning, and used media queries where the guided one used container queries. LogRocket sells its own AI product, and the article says Chrome calls the project an "early preview".
Install paths and defaults
The docs list Claude Code, Codex, Gemini CLI, GitHub CLI, Antigravity, Copilot CLI, Goose, Grok, Kimi Code and JetBrains WebStorm as supported, plus npx skills. The docs say most of the non-CLI options do not support automatic updates.
| Agent | Install command |
|---|---|
| Any, with auto-updates | npx modern-web-guidance@latest install |
| Claude Code | claude plugin install modern-web-guidance@googlechrome |
| Gemini CLI | gemini extensions install https://github.com/GoogleChrome/modern-web-guidance --auto-update |
| GitHub CLI | gh skill install GoogleChrome/modern-web-guidance |
Without a Baseline target set in AGENTS.md or CLAUDE.md, guidance defaults to Baseline Widely available. Features that ship in only one browser engine need set-allow-volatile-features, and origin-trial features need set-allow-origin-trials. Stars measure interest, not adoption, and neither Chrome page publishes install counts.
Two other developer-tooling stories on this site: Makepad's Relax runs Rust without rustc, and Rust 1.100 stabilizes the never type. The Chrome docs give no release date, and the README links an evaluation dashboard whose results this report did not read.
Sources
More in A.I.
- 01Qwen3.8-27B's Model Card: Apache 2.0, a 1M-Token Ceiling and Scores Run Through Claude CodeAlibaba's 27-billion-parameter dense model beats Claude Opus 4.6 Max on SWE-bench Pro by its own count, on a task set Alibaba modified.
- 02Apple's Harbor App Would Cluster Thunderbolt 5 Macs for Local AI, per a LeakerThe only source for the unreleased app is leaked text attributed to Aaron Perris; Apple has announced nothing, though the clustering layer underneath it is documented.
- 03Nous Research's Arete-27B Cuts Propaganda Answers From 22% to 0.9% on Its Own TestThe Qwen3.8-27B fine-tune is Apache-2.0 and open on Hugging Face, but the evaluation behind the headline number was built by its trainers.
- 04Edge0-145B's Model Card Names Qwen as Its Base, and the Launch Post Never DoesThe 145-billion-parameter open-weights release says it runs on a laptop, but its licence is two documents and its benchmarks are all Edge0's own.