Modern Web Guidance: Claude Code's 57% to 88% in Google's Eval

Claude Code passed 57% of Google's web checks without this skill pack. With it, 88%. Those are Google's own eval results for Opus 5.5. The pack is free, and its new scanner looks for older patterns in existing code.

What Chrome released on November 10

Google's Chrome team released Modern Web Guidance 1.0 on November 10, 2026. Firefox and Edge joined the project, which is stewarded by the Chrome, Edge, Firefox and MDN teams alongside the web development community.

It's a set of coding-agent skills with web expertise, best practices and browser compatibility data. The release includes 203 use-case guides covering 250 web platform features.

Why agents need the guidance

The project's README says agents often default to older patterns because their training data contains so much legacy code. Chrome's team gives a concrete example: an agent can write 100 lines of custom JavaScript for something modern HTML and CSS handle in a few lines.

That's the problem the guides target. They point the agent toward native browser features when those fit the job.

Google's eval results

In Google's own November 8 eval, the share of checks passed changed as follows:

Agent and model	Without the skill	With the skill
Claude Code, Opus 5.5	57%	88%
Codex CLI, gpt-6.1-sol	54%	82%
Antigravity CLI, gemini-4-argon	58%	94%

Google uses realistic developer prompts against full base applications. Playwright graders check source patterns, computed styles, runtime DOM behavior and accessibility states. These are Google's measurements on its tasks. They don't promise the same improvement on your code.

Scan an existing repo

New in 1.0, legacy-scan checks HTML, CSS, JavaScript and package.json for a limited set of documented legacy patterns. Examples include JavaScript scrollHeight textarea auto-resize hooks and premature :invalid styling on untouched inputs. It maps each finding to a guide with the modern replacement.

The agent workflow has 4 steps: scan, verify, report, modernize. Verification comes before changing the code.

Google also recommends pairing the guidance with Chrome DevTools for agents to run and verify the result.

MetaMask's review bot

Per Google's post, MetaMask runs the guidance as a review bot on every pull request to its open source browser extension. In its first five weeks, it reviewed about 480 PRs and commented on 64. Engineers fixed 43 findings, mostly accessibility, HTML semantics and forms issues. Google reports a median of 15 minutes from comment to fix.

That's a customer story published by Google. Your own repo may show different results.

Install it

It works with Claude Code, Codex, Gemini CLI, Copilot CLI and Antigravity. Install with:

npx modern-web-guidance@latest install

Then scan a repo:

npx modern-web-guidance@latest legacy-scan .

The code is Apache 2.0, and the guides are CC-BY 4.0. The pack is free. Use it with the agent you already pay for.

The catch

The scanner covers roughly 20% of the guides so far. Static analysis checks a limited set of patterns, so a clean scan doesn't cover the whole pack.

The evals are Google's own. Anonymous usage telemetry is also enabled by default: install counts, calling agent name, retrieved guide IDs and agent-generated search queries. Raw user prompts aren't collected.

To opt out, set DISABLE_TELEMETRY=1. For example:

DISABLE_TELEMETRY=1 npx modern-web-guidance@latest legacy-scan .

The builder angle (my opinion, no income guaranteed)

I'd offer a fixed-scope front-end modernization audit to an agency or small business with an older site:

• Run the scan on a copy of the client's repo.
• Verify each finding against its linked guide and the site's behavior.
• Deliver a report listing native HTML and CSS replacements, with accessibility fixes flagged.
• Quote the cleanup as a separate job.

The report is the audit deliverable. My suggestion is to agree on that scope before proposing implementation work.