Browser vendors ship a skill pack for coding agents, with evals attached

Modern Web Guidance is a set of agent skills and a local CLI. The agent searches about 200 hand-written guides on current HTML, CSS and browser APIs and loads the matching guide into its context. The 1.0 release, announced by Philip Walton on the Chrome for Developers blog, adds a static legacy scan that flags old patterns to modernize, plus a public eval dashboard. The README states the problem directly: "Coding agents often default to older patterns because LLM training data contains vast amounts of legacy code."

The evals give agents realistic tasks on real web apps. Playwright and AST assertions grade the results, and each grader is calibrated to pass on a known-good solution and fail on the unmodified app. In the November 8 snapshot, on about 215 tasks, Claude Code on Opus 5.5 went from 57% to 88% of assertions passed with the guides. Codex CLI on GPT-6.1 Sol went from 54% to 82%, and Antigravity CLI on Gemini 4 Argon from 58% to 94%. These are the maintainers' own figures, and we have not rerun them.

Paul Irish of the Chrome team posted the release on Hacker News (144 points at about 16:00 EST). He wrote that the team prunes a guide when evals show models already know its material, and that the legacy scan covers about 20% of what the guides address. One commenter reported that the scan found only two issues in their code; Irish replied that it is tuned to avoid false positives. Telemetry is on by default and can be turned off with DISABLE_TELEMETRY=1.

Why it matters: This skill comes with an uplift measurement rather than only a description of what it should do, and the method is one a team can copy for its own skills: graders that must fail on the base app and pass on a reference solution, then guided and unguided runs compared. If agents write your frontend code, running the legacy scan on a repo you already know is a cheap first check of whether the guides find anything.