TLDR;
Philip Walton introduces Modern Web Guidance, a tool developed by the Chrome team to address the shortcomings of AI coding agents in modern web development. AI agents often recommend legacy solutions, lack awareness of new web features, and make outdated browser support recommendations. Modern Web Guidance provides expert-vetted skills and best practices, ensuring AI agents can leverage the latest web platform features effectively and make informed decisions about browser compatibility. The tool includes high-level guides for web disciplines and specific guides for new web features, optimized for AI consumption and cross-browser support.
- AI coding agents often recommend legacy solutions and lack awareness of new web features.
- Modern Web Guidance provides expert-vetted skills and best practices for AI agents.
- The tool includes high-level guides and specific guides for new web features, optimized for AI consumption and cross-browser support.
Intro [0:00]
Philip Walton introduces a presentation about the current state of AI coding agents in web development. He demonstrates two identical demos generated by AI, one with standard configuration and the other with an additional tool. The key difference is that the standard version loads 172KB of JavaScript, while the enhanced version loads less than 3KB and uses no JavaScript, highlighting a significant inefficiency in AI-generated code. This comparison raises the question of why AI agents recommend JavaScript-heavy solutions when more efficient, JavaScript-free alternatives exist.
Problems with AI Coding Agents [3:23]
AI coding agents often recommend legacy patterns despite the availability of better, more modern alternatives. These models have knowledge gaps, lacking awareness or having incorrect information about the latest web platform features due to knowledge cutoff dates. For example, an AI model used the incorrect attribute interesttarget for the interest invokers API, which was changed to interestfor before the API shipped in Chrome. AI models also make conservative, one-size-fits-all browser support decisions, failing to leverage compatibility data from initiatives like Baseline.
Introducing Modern Web Guidance [5:47]
Chrome introduces Modern Web Guidance, a solution designed to address the problems with AI coding agents. This comprehensive set of skills embeds Chrome's web platform expertise directly into coding agents, working with any AI coding platform. It can be installed using the command npx modern-web-guidance install or directly from within supported IDEs and coding agents. Modern Web Guidance is a long-term investment by the Chrome team to ensure developers can build effectively for the modern web.
Components and Functionality of Modern Web Guidance [7:18]
Modern Web Guidance includes high-level guides for broad web disciplines like performance, accessibility, and security, as well as specific low-level guides for recently added web features. These guides are created by the Chrome team and external subject matter experts. When a user enters a prompt, the AI performs a semantic search within the Modern Web Guidance package to find matching guides, which are then added to the context window. This approach differs from traditional skills packs by using a single entry point skill with access to a CLI for searching the extensive guide database.
Guide Structure and Optimization [9:21]
Modern Web Guidance features high-level discipline skills and low-level feature guides. The feature guides focus on distinct use cases that new web features unlock, rather than documenting the features themselves. For example, a use case for the View Transitions API is "Visually connect persisting elements across different page states," without explicitly mentioning the API. These guides are optimized for AI consumption, being concise and containing only the necessary information. Each guide includes recommendations and fallback strategies to ensure cross-browser compatibility.
Cross-Browser Support and Baseline Integration [11:40]
Each guide documents the ideal implementation using the most modern solution available, without initial consideration for browser support. If any features are not Baseline widely available, a cross-browser fallback recommendation is included. Developers can specify their project's browser support requirements in an AGENTS.md file, allowing the agent to tailor the implementation accordingly. If no specific browser support is declared, the agent assumes Baseline widely available. The agent uses browser compatibility data to determine whether to use a fallback based on the specified requirements.
Evals and Future Plans [14:31]
Modern Web Guidance includes a rigorous evaluation suite with 554 distinct assertions to ensure the effectiveness of each feature. This suite is run daily against popular AI coding models like Gemini 3.1, Claude Opus 4.7, and GPT 5.5. The guided pass rate (with Modern Web Guidance) is compared to the unguided pass rate to measure improvement. The plan is to continuously monitor these evals and remove use cases once the unguided version performs consistently well. The goal is to cover every new web platform feature that ships in Chrome by the time it reaches Chrome stable, and to collaborate with other browser vendors for broader coverage.
Demo: Enhancing a Mobile Email App [19:20]
Philip demonstrates Modern Web Guidance by enhancing a mobile email application. He uses Antigravity, running two versions: one guided with Modern Web Guidance and one unguided. The initial problem is that the navigation drawer doesn't swipe to close, a common feature in native apps but lacking in many mobile websites. He prompts both versions of Antigravity to fix this. The unguided version implements a basic swipe-to-close, but the guided version provides a smoother, more intuitive experience using scroll-driven animations and scroll snap points.
Adding Drilldown and Swipe Actions [26:53]
The next prompt asks the AI to add a message view drilldown, allowing users to click an email in the inbox and see the full message with a Back button. The unguided version creates a functional drilldown, but lacks smooth transitions. The guided version, however, implements a smooth transition using the stack drilldown module. Finally, the AI is prompted to add swipe actions for archiving or snoozing messages. The unguided version again provides a basic implementation, while the guided version delivers a smooth, Gmail-like experience with messages pinned to the user's finger during the swipe.
Conclusion and Call to Action [33:32]
Chrome is committed to improving the web by making it easier for developers to use new features. Modern Web Guidance is designed to help developers build better experiences by providing expert-vetted recommendations and ensuring AI coding agents leverage the latest web platform features effectively. The Chrome team encourages developers to try out the early preview and provide feedback to help improve the tool.
