AI Resource Library Evaluated methodologies, workflows, and tools
← All resources

Modern Web Guidance: on-demand web platform expertise for coding agents

catalog · https://developer.chrome.com/docs/modern-web-guidance · by Google Chrome team · Evaluated 25 May 2026
claude-codeskillsweb-platformcssfrontendagent-contextpreview

What it proposes

An agent skill (a SKILL.md plus a CLI) that injects curated modern web-platform knowledge into a coding agent’s context window on demand, rather than relying on whatever the model absorbed from training data. The premise: LLM training corpora are dominated by legacy code, so agents default to bloated, dated patterns even for tasks that now have native platform solutions, and even when a model knows an API exists it lacks the density of production-ready implementation patterns.

The mechanism is a retrieval CLI with three commands: install (an interactive wizard that places the SKILL.md), search "<query>" (find relevant guides by natural-language description), and retrieve "<guide-id>" (pull a specific guide into context). It ships roughly 102 feature guides across CSS and layout (container queries, subgrid, oklch, text-wrap, line-height trimming), HTML and DOM, UX (View Transitions, entry/exit animations), performance (instant preloading, INP diagnostics, scheduler.yield), forms and UI (Anchor Positioning, Popover API, <dialog>, :user-invalid, field-sizing), accessibility (error announcements, focus management), and built-in browser AI APIs. The guides are deliberately token-efficient: evals prune content the models already know reliably, so only the high-value, hard-to-recall patterns are loaded. It installs via its own CLI, via Vercel’s Agent Skills, or as a plugin/marketplace entry for Claude Code, Copilot CLI, and Antigravity CLI, and integrates with Chrome DevTools for agents for real-time auditing. Supported by the Chrome and Edge teams; currently a preview release.

Best used when

  • The agent regularly builds or maintains interactive web UI: components (accordions, tab bars, dashboard cards), dialogs and tooltips, forms, animations, or anything touching layout and styling.
  • A codebase is migrating off legacy patterns (jQuery modals to <dialog>, custom tooltip positioning to Popover API plus Anchor Positioning) and you want the agent to reach for native platform features instead of reinventing them in JavaScript.
  • Performance and accessibility of a real front end matter (INP/LCP work, focus management, accessible error announcements) and you want expert-curated, current guidance loaded just-in-time rather than trusting the model’s stale defaults.
  • You want the knowledge available across multiple agent CLIs from one source, loaded only when a task actually needs it, keeping the context budget lean.

Poor fit when

  • The project’s centre of gravity is knowledge work, writing, or content curation rather than building feature-rich web applications. The guidance targets dialogs, popovers, anchor positioning, view transitions, and similar interactive primitives that such projects rarely implement, so the skill sits idle.
  • The only web surface is a static content site (a blog, a documentation site, an Astro site rendering markdown). These benefit marginally at most: they style and lay out content but seldom build the interactive, stateful UI the bulk of the guides address, and a static-site generator’s own conventions usually dictate the limited CSS involved. Installing it adds an agent skill that almost never fires.
  • The team has no recurring front-end UI workload at all. The payoff scales with how much interactive web code the agent writes; near zero such work means near zero return, even though the tool itself is free and local.

Verdict

Catalog. This is a genuinely high-quality, well-targeted tool: free, local (npx/CLI), backed by the browser vendors who define the platform, and built on a sound insight that agents default to dated patterns and benefit from token-efficient, eval-pruned, just-in-time guidance. For any project whose agent routinely builds modern web UIs, adopt is the obvious call. The reason it lands at catalog rather than adopt here is scope, not quality or cost: a workflow centred on vault/knowledge work, writing, and static content sites rarely produces the interactive web features the guides specialise in, so the skill would seldom activate. The preview-release status is a minor secondary caveat (surface area and guide IDs may shift), but it does not drive the verdict; even at 1.0 the applicability gap would remain. Worth knowing about and worth adopting immediately should a front-end-heavy project enter the picture; not worth installing into a project that doesn’t build interactive web UI. The transferable idea, independent of web development, is the retrieval pattern itself: an eval-pruned, search-and-retrieve knowledge skill that loads only what the model demonstrably doesn’t already know, which generalises to any domain where an agent’s training-data defaults are stale.