Source: GoogleChrome/modern-web-guidance README (Apache-2.0); stewarded by the Chrome, Edge, Firefox and MDN teams. Surfaced by André Lindenberg.
The problem
Coding agents learned the web from years of legacy code. They reach for JavaScript and polyfills for things the platform now does natively (dialogs and pop-overs, top-layer animation, scroll-driven effects), producing heavier, less accessible pages. Knowing an API exists is not the same as knowing the modern implementation pattern.
What it is
A skill set plus CLI that puts curated, browser-compatibility-aware guidance into the agent's context before it writes UI code:
searchfinds relevant guides for a task (the post cites 209 guides);retrieveloads one by ID.legacy-scanscans a codebase for legacy patterns to modernize.Guides cover modern browser APIs, performance and accessibility, and lightweight fallbacks instead of heavy polyfills.
Guides are pruned with evals so they are token-efficient: content models already know is dropped. An eval dashboard is public.
Installs via
npx modern-web-guidance installinto common agents.
Why it matters
It is a model for vendor-maintained skills: the people who ship the platform publish the guidance agents should read, with compatibility data attached.
Retrieval at write time beats retraining for fast-moving APIs.
legacy-scangives a measurable target: count legacy patterns before and after an agent-driven refactor.
Caveats
Per the post, search queries go to Google unless telemetry is disabled; check the setting in regulated codebases.
Guidance favors modern APIs; teams with strict older-browser support still need their own baseline policy.
A skill only helps if the agent actually calls it; measure that, as with any skill.
Review checklist
What browser baseline do we target, and does the guidance respect it?
Is telemetry configured as our policy requires?
Do we measure legacy-pattern counts in agent-written UI code?
Does our agent call the skill on UI tasks, or ignore it?
Related entries
R091 (serving skills from a server), R064 (measuring whether a skill helps), R004, R044.