Synergised Consulting
Technical asset

Google's Modern Web Guidance 1.0 Fixes What Your AI Coder Never Learnt

Synergised Consulting Ltd

8 min read
Title card: Modern Web Guidance 1.0, with the unguided-to-guided evaluation pass rate rising from 51 to 91 per cent

Last Updated: 18 November 2026

Your AI coding agent writes web code a senior front-end developer would send back, because it repeats the most common patterns in its training data, and most web code on the internet is a decade old. Google Chrome's Modern Web Guidance 1.0 is a free, open-source skill pack that puts current web platform guidance in front of the agent at the moment it writes. On Google's public evaluation dashboard, unguided frontier agents passed about 51 per cent of modern web checks; with the skill installed, the same agents passed about 91 per cent [1].

What Modern Web Guidance Actually Is

Modern Web Guidance is a set of skills that embeds web platform expertise, best practices and browser compatibility data directly into AI coding agents, stewarded jointly by the Google Chrome, Microsoft Edge, Mozilla Firefox and MDN teams and released under an Apache-2.0 licence [2]. It shipped as an early preview at Google I/O in May 2026, and the stable 1.0 release was announced on 10 November 2026 [3]. Version 1.0 carries 203 use-case guides covering 250 web platform features, and installs into agents including Claude Code, Codex CLI, Cursor, GitHub Copilot and Antigravity with a single command [4].

The design decision that makes it practical is retrieval, not loading. At rest the agent sees only a skill description of roughly 205 tokens; when a front-end task comes up, it searches the corpus and pulls exactly one guide of roughly 800 tokens into context [2]. A guide database of this size dumped into the context window every turn would be wasteful on cost and diluting on attention, so nothing enters context until the agent needs it.

What the Evaluations Actually Measure

The headline numbers come from Google's own evaluation harness, and they are stronger evidence than most vendor benchmarks because of how the grading works [1]. Each task pairs a starter web application with a realistic prompt such as "add search-term highlighting", deliberately avoiding the names of the APIs involved so the agent has to recognise which guidance applies. A Playwright grader then opens the finished app in a real browser and checks what the code does, not what it looks like: rendered behaviour, computed styles, DOM state and accessibility semantics [1].

The calibration runs in both directions. A golden solution must score 100 per cent and the untouched base application must score zero, so a grader that passes everything or fails everything is rejected and rewritten. Agents run in sandboxed environments so they cannot read the grader and reverse-engineer the expected output [4]. Results are published on a public dashboard showing per-agent and per-guide pass rates, with nightly jobs re-running the comparison so the numbers stay current rather than frozen at launch [1].

For a business owner, the useful reading of a 51-to-91 per cent swing is not about the tool. It is evidence of how much of an agent's front-end output is governed by what sits in its context when it writes. The same model, on the same tasks, produces materially different code depending on whether current guidance is present. That is a lever you control.

How Baseline Decides When Code Needs a Fallback

The mechanism underneath the guidance is Baseline, the web platform's compatibility signalling system, which the skill uses to decide whether a feature can be used bare or must ship with a fallback [2]. Baseline has three statuses: Widely available means every core browser has supported the feature for at least 30 months; Newly available means it works everywhere core but not for long enough; Limited availability means at least one core browser does not support it [5].

Worked example. Consider a confirmation prompt. The modern platform answer is the native dialogue element the HTML standard now provides, paired with an attribute that adds light dismissal: the browser gives you the focus trap, the Escape key handling, the stacking and the accessibility semantics without you writing them. But that attribute is Limited availability: Chrome and Edge shipped it in 2025, Firefox later the same year, and Safari had not shipped it at the time of writing. The guide does not pretend the attribute is safe to use bare. It hands the agent the modern markup plus a small click listener covering the browsers that lack the attribute. The result is clean code on current browsers and correct behaviour on Safari, without the developer knowing which browser shipped what and when.

The fallback ladder has four rungs, cheapest first: progressive enhancement where the feature is additive and nothing breaks without it; a CSS @supports block providing a simpler appearance when the feature is missing; a small amount of custom JavaScript where the behaviour genuinely needs it; and, last, a polyfill loaded conditionally so it never reaches a browser that does not need it [2]. The rules are also project-configurable: the recommendations draw on your project's browser targets, so a team supporting managed government browsers can hold the line at older Baseline levels while an internal dashboard on a known Chrome fleet can use newer features without ceremony [1].

What This Means If Agents Build Your Software

This is a product launch with one very direct before/after, and it is worth stating as an illustration rather than a case study. Before: an agent asked for a modal writes a stack of divs, a hand-rolled focus trap, a keydown listener, a backdrop-click handler and a body scroll lock, and misses the reduced-motion preference entirely. After, with the skill installed: the same request produces the native element, a small fallback for the one browser that lacks the attribute, and the platform's accessibility semantics included by default. The diff is smaller, the review is faster, and fewer defects reach a human.

This is also a pattern we run daily from the other side. Our own agents work the same way this skill pack assumes: a narrow skill description at rest, one focused document pulled in when the task needs it, current guidance instead of stale memory. In our discovery conversations with owners, the question is rarely whether an agent can do the task; it is whether the agent will do it the way the business actually needs, every time. The answer in both cases is the same mechanism: put the current, specific answer in front of the agent at the moment it works, rather than hoping the model's training data covers it.

The compounding effect is operational, and it stays on the operational ledger. An agent that writes less code produces smaller diffs, and smaller diffs are cheaper for a person to review. For a founder who has just acquired a business and is putting agents into its workflows, that is the difference between a toolchain that scales review capacity and one that silently consumes it.

Where It Falls Short

Modern Web Guidance is a retrieval layer and a guidance policy, not a guarantee and not an audit, and three specific limits matter [1].

First, there is no way to guarantee a guide is retrieved and applied on a given task. Google says this plainly [1]. An agent can miss the lookup, or recognise the task incompletely, and the failure is silent: it falls back to the pattern in its training data, which is exactly the behaviour the skill exists to correct.

Second, Baseline is a compatibility baseline, not an accessibility baseline. A feature can be Widely available and still be poorly exposed to screen readers, and a Baseline-ready pattern can still fail WCAG conformance. Adrian Roselli [6], an independent accessibility researcher, tested the May preview and found an accordion component that added unrequested features, failed its animation in Firefox despite the fallback promise, and was not WCAG-conformant, while conceding that output is almost certainly better than doing nothing. The project has since added a dedicated accessibility review step for user-facing guides in its 1.0 contribution process [3], but no independent 1.0 retest had been published at the time of writing, and the structural criticism stands: a green compatibility badge never substitutes for an accessibility audit.

Third, it is not for every team. If your agents write only back-end services, internal scripts or data pipelines, a web platform skill pack has nothing to correct, and adopting it adds ceremony for no return. Google's own positioning is front-end work [1].

A One-Day Adoption Path

The sensible adoption path is a pilot on one repository, not a policy change. Install the skill into one agent using the project's own one-line commands [1], run its new modernisation workflow, which scans an existing codebase for safe modernisation opportunities [3], and compare a handful of pull requests written with and without it. Adopt it if your agents write front-end code; skip it if they do not; and in either case treat it as a reducer of defects reaching review, never as a replacement for review.

One governance note belongs in the onboarding checklist rather than the docs: the pack is stewarded by the browser vendors and ships fast, so decide deliberately whether you pin a version or track latest, and record that decision next to your Baseline target. An agent whose guidance changes between sessions without anyone deciding it should change is an agent nobody is governing.

Sources

  1. [1] Google Chrome for Developers, Modern Web Guidance documentation:
  2. [2] GoogleChrome/modern-web-guidance, GitHub repository:
  3. [3] Modern Web Guidance v1.0 launch announcement, Chrome for Developers blog, 10 November 2026:
  4. [4] Modern Web Guidance v1.0.0 release notes:
  5. [5] Baseline web platform compatibility definitions, web-platform-dx working group:
  6. [6] Background reading (qualitative guidance, no figures): Adrian Roselli, independent accessibility critique of the Modern Web Guidance preview: