Modern Web Guidance: Chrome's Agent Skill Against 2019 JavaScript

Last updated: November 11, 2026

Chrome, Edge, Firefox and MDN have shipped Modern Web Guidance 1.0, a skill and CLI that tells coding agents to reach for native platform features instead of the <div> plus JavaScript patterns that dominate their training data. If you searched for modern web guidance, this is the short version: what it installs, what the team claims it measures, and the one design question that decides whether you should switch it on.

What the primary source says#

The project README describes Modern Web Guidance as "a set of skills that embed web platform expertise, best practices, and browser compatibility data directly into your coding agents." Its stated problem is that training data contains a lot of legacy code, so agents "generate bloated JavaScript for tasks that now have native, high-performance web platform solutions" and, even when a model knows an API exists, it "lacks the density of real-world, modern implementation patterns."

The mechanism is a CLI the skill calls, which returns targeted guides into the context window instead of loading everything. The README documents three commands you can run 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 .

npx modern-web-guidance@latest install runs an interactive wizard. The repository carries plugin manifests for Claude, Codex, Cursor and Gemini, among others, so it is built to sit in whatever agent you already use rather than to replace it.

The README lists 250 modern web features, 140 of them in CSS and layout, across user experience, layout, performance, forms, accessibility and built-in browser AI. The guides are written to be token-efficient, and the team says it runs evals so it can prune content models already know.

The numbers are vendor-reported#

In the Show HN post, Paul Irish of the Chrome team gives the headline figures. All of these are vendor-reported, not independently reproduced:

• 200+ curated guides, about 1.7 MB in total.
• An eval of 200 tasks with roughly 1,700 Playwright and AST-based assertions, run against Codex, Claude and Antigravity. Agents went from 55% pass unguided to 92% guided.
• A deterministic legacy-scan trained on 30,000 GitHub repos, with 92% average precision on held-out repos.
• The retrieval layer shrank from a 500 MB native stack to a roughly 7 MB pure-JS package with search around 10 ms on CPU.

The eval is the interesting part, because it is the opposite of a prompt-lore skill. It gives you a measured before and after and a mechanism for deleting guides the model no longer needs. That matches the exit-criteria argument in Agent Skills Need Exit Criteria: a skill should be able to prove it earns its tokens. Treat 55% to 92% as a ceiling for a benchmark built by the authors, not a number your own repo will hit.

The angle: a skill is a policy, and this one has an opinion#

The most useful reading of this launch is not "Chrome made a skill." It is that a vendor is now publishing agent behavior as a versioned, evaluated package, the pattern covered in Agent Skills Are Becoming Package Managers and anticipated by Google Skills. Skills are where platform owners can correct what models learned from old code.

That is also where the risk sits. A skill that says what to prefer is a policy, and policies encode a stance on browser support. If you use Claude agents or skills for front-end work, the question is whether the stance matches your audience.

What people are actually saying#

The Hacker News thread reached about 300 points, and the argument is concentrated on one point.

The critique: one commenter listed CSS features the guides discuss that are stable in Chrome but missing from Firefox and Safari, and another said that as a target user they would expect a Widely Available default with an opt-in for newer features. A third noted that some guide wording reads as an instruction ("DO use fetchLater()") even if it is meant as documentation.

The reply: Chrome's Philip Walton said the guides document the features and include a browser support table for any API that is not Baseline Widely available, usually with a cross-browser fallback, so the agent decides from your support requirements. He also said the goal is not to push Chrome-only APIs, and described the test as what an expert consultant would recommend, including when to wait. On fetchLater() he said it is being added to WebKit, a claim from the thread rather than something this post verified.

Both sides can be right. The documentation is careful, but a default that leans modern is still a default, and the burden shifts to you to state your support floor.

What to do Monday#

1. Run the read-only scan first: npx modern-web-guidance@latest legacy-scan . in a front-end repo. It tells you how much old-pattern code you have before you change any agent behavior.
2. Try one search for something your agents get wrong, such as a modal or an auto-growing textarea, and read the retrieved guide yourself.
3. If you install it, put your browser support floor in your agent instructions. For example: "Support only Baseline Widely available features unless a task says otherwise." That addresses the main HN objection directly.
4. Re-run your own tasks with and without the skill for a week. The authors' 92% is their benchmark, not yours.

FAQ#

What is Modern Web Guidance? An open-source set of guides, shipped as an agent skill plus a CLI, stewarded by the Chrome, Edge, Firefox and MDN teams.

How do I try it without installing? Use the search and legacy-scan commands above.

Will it make my agent use features Safari or Firefox lack? Only if you let it. The guides include support tables, so set your support floor explicitly.

Sources#

Continue Reading#

• Agent Skills Need Exit Criteria - how to tell whether a skill earns its tokens
• Best Claude Code Skills and Plugins - where to find and choose skills
• Claude Agents vs Skills - which one you actually need
• Agent Skills Are Becoming Package Managers - versioned, governed skill distribution
• Google Skills - Google's earlier skills playbook