agentic-browsing-auditor
Verify and improve a website, current page, or frontend repository for Lighthouse Agentic Browsing readiness. Use when asked to audit, raise, improve, or debug agentic browsing level, WebMCP readiness, llms.txt discoverability, agent-accessible forms, accessibility-tree quality for agents, or layout stability for AI/browser agents.
- License 许可
- MIT
- 最近更新
- 2026-11-06
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- agentic-browsing-auditor
- description
- Verify and improve a website, current page, or frontend repository for Lighthouse Agentic Browsing readiness. Use when asked to audit, raise, improve, or debug agentic browsing level, WebMCP readiness, llms.txt discoverability, agent-accessible forms, accessibility-tree quality for agents, or layout stability for AI/browser agents.
02 资源文件 3
- 参考文档agents/openai.yaml257 B
- 参考文档README.md3.4 KB
- 参考文档references/chrome-agentic-browsing-checklist.md4.1 KB
03 SKILL.md 正文
Agentic Browsing Auditor
Core Idea
Optimize the page for deterministic machine interaction, not only visual browsing. Treat Chrome's experimental Lighthouse Agentic Browsing category as a readiness checklist: expose structured tools, keep forms self-describing, make the accessibility tree complete, reduce layout shifts, and publish a root llms.txt.
Read references/chrome-agentic-browsing-checklist.md when you need the detailed scoring notes or implementation heuristics.
Workflow
- Identify the target: a live page, current browser page, local dev server route, static HTML file, or repository.
- If the page is live or runnable, inspect it in a browser first:
- Open the page in Chrome or the in-app browser.
- Run Lighthouse Agentic Browsing audits when available.
- If the Agentic Browsing category is unavailable, do not invent a substitute score; use browser inspection and repository evidence instead.
- Capture the failing audit names and exact nodes/forms/tools before editing.
- Inspect visible interaction flows and accessibility names for the primary tasks.
- If working from a repository, inspect the implementation directly:
- Search for page routes, templates, components, forms, custom buttons/links, modal/dialog components, and file-serving routes.
- Search for WebMCP declarations:
toolname,tooldescription,toolparamdescription, andnavigator.modelContext.registerTool. - Search for discoverability files or routes:
llms.txt,robots.txt, sitemap generation, docs indexes, API references, and public route manifests. - Search for layout-shift risks: images/media/iframes without stable dimensions, late-loaded widgets, hydration-only content, ad slots, banners, toast regions, and skeletons.
- Search for accessibility risks: icon-only buttons, clickable non-buttons, custom selects/comboboxes,
aria-hidden,role,tabindex, and controls without labels.
- Fix the highest-leverage issues in this order:
- Broken WebMCP schemas: partial
toolname/tooldescription, required controls withoutname, malformed imperative registrations. - Missing WebMCP coverage on primary task forms.
- Missing programmatic names for buttons, links, inputs, and other interactive controls.
- Layout instability risks, especially media without dimensions and injected content above task controls.
- Missing or weak
llms.txtas a discoverability improvement unless a configured route/server error makes it a confirmed failure.
- Broken WebMCP schemas: partial
- Re-run live checks when available. If no live page is available, report repository findings as static evidence and mark browser/Lighthouse verification as not run.
Repository Inspection
Use source inspection to find the code paths that define the page, then judge each issue from the surrounding framework patterns. Do not rely on broad regex matches alone; trace components through props, wrappers, and generated routes before calling something a failure.
Useful search targets:
toolname,tooldescription,toolparamdescription,registerTool,modelContext<form,required,name=,type="submit", form actions, server actions, validation schemas<button,role="button",onclick,on:click,@click,tabindex, icon button components<input,<select,<textarea,aria-label,aria-labelledby,<label<img, image components,iframe,video,aspect-ratio,width,height, lazy-loaded mediallms.txt,robots.txt,sitemap, public/static asset directories, route handlers
Classify findings:
Failure: the source clearly creates an invalid WebMCP schema, unnamed required field, unnamed interactive control, inaccessible task control, broken configuredllms.txtroute, or server error for an expected public discovery file.Warning: the source suggests a risk but needs browser rendering, framework resolution, or runtime state to confirm.Opportunity: the page works, but a clear WebMCP tool, new or betterllms.txt, or more stable layout would make agent use easier.
Improvement Rules
WebMCP tools
Prefer declarative WebMCP for normal HTML forms. Add both attributes together:
<form toolname="newsletter_signup" tooldescription="Subscribes the user to the newsletter">
<label for="email">Email</label>
<input id="email" name="email" type="email" toolparamdescription="The user's email address" required>
<button type="submit">Sign up</button>
</form>
Use imperative navigator.modelContext.registerTool only for dynamic tools, non-form actions, stateful flows, or actions that need custom execution. Keep each tool single-purpose, action-oriented, and available only in page states where it can actually run.
Forms and schemas
Every tool form must have both toolname and tooldescription. Every required form field inside a tool form must have a stable, unique name. Prefer explicit toolparamdescription for fields whose purpose is not obvious from an associated label.
Accessibility tree
Agents rely on the accessibility tree as their page model. Make every interactive element discoverable by role and name: use semantic HTML, visible labels, aria-label/aria-labelledby only when needed, valid ARIA relationships, and no focusable or clickable controls hidden from assistive technology.
llms.txt
Add llms.txt at the served domain root when it would improve agent discoverability. Treat absence as an opportunity by default, not a failure. Treat it as a failure only when the site already advertises or implements the route and it returns a server error or broken content. Keep it concise Markdown with the site's purpose, canonical key pages, task entry points, docs/API references, and any constraints agents need to know. Avoid dumping entire docs or private/internal-only locations.
Layout stability
Reserve space for images, video, embeds, ads, async widgets, and late content. Set dimensions or aspect ratios, avoid inserting content above active controls after load, and make skeleton/loading states size-stable.
Reporting
Lead with confirmed failures and the user-visible impact for agents. Include file/line references for repository findings. Separate:
Failures: likely to fail Lighthouse or block agents.Warnings: may reduce reliability or need manual confirmation.Improvements made: exact files changed.Verification: live Lighthouse/browser checks, repository inspection, and any remaining gaps.
04 同仓库更多 Skills
motion-design
Plan, build, and review short motion-design films for SaaS and app marketing: launch videos, product films, promo and social ad clips, app trailers, and landing-page loops of roughly 6–45 seconds. Develops the brief, promise, storyline, rendered key frames, timed beat sheet, and sound plan first, has them critiqued with fresh eyes, and gets sign-off before any animation, because iterating on a plan is far cheaper than iterating on renders. Then builds the film (following the project's motion studio when one exists, otherwise a code-rendered timeline) and reviews cuts with a scored rubric and prioritized fixes. Use whenever the user wants to make, script, storyboard, animate, render, critique, or fix a product video, motion graphics ad, launch clip, or app promo, even if they only say 'make a video for my app' or share a cut for feedback. Not for UI micro-interactions inside the product itself.
mise
Set up, migrate, or maintain a project on the team's mise stack: mise tools and tasks under one shared task-naming convention, deployment environments as mise config environments (mise.<env>.toml), uv for Python, bun for JavaScript and TypeScript, Terraform in terraform/, and optional fnox secrets (Proton Pass locally, environment variables in CI and on deployed hosts). Invoke it to replace a Makefile, justfile, or package scripts, add or rename tasks, add secrets or environments, or bootstrap a repository.
grafana
Set up lean application monitoring in Grafana Cloud for a solo maintainer running several web apps, APIs, mobile backends, or jobs on GCP or AWS. Use when adding metrics or monitoring to an app; sending OpenTelemetry (OTLP) or Prometheus metrics to Grafana Cloud; querying Cloud Run, Lambda, Cloud Monitoring, or CloudWatch metrics from Grafana; building or fixing app and portfolio dashboards; or keeping active series within the free tier. Also covers client-side health signals for mobile and browser apps and whether a first alert is worth adding.
google-keyword-research
Research Google keywords for SEO and paid search using Google Ads planning data (Keyword Planner). Use for keyword ideas, search volume, "how many people search for" questions, CPC and bid estimates, competition, seasonality, or a Search Viability Loop that prices commercial keyword clusters against contribution-based acquisition ceilings and read-only forecasts before proposing a paid test.
frontend-design
Design, implement, and refine polished web interfaces: new screens, components, dashboards, landing pages, and visual redesigns. Use when UI quality, layout, typography, responsive behavior, or interaction polish is part of the task, including requests to make an interface look better. Reuses or establishes a design system and verifies the rendered result. Not for backend-only work or logic changes with no visual impact.
apple-ads-keyword-research
Read Apple Ads (formerly Apple Search Ads) campaign data and research App Store keywords through the Apple Ads Platform API. Use for campaign, ad group, keyword, or search term performance (spend, taps, installs, CPT, CPI), Apple's keyword and phrase suggestions, relative App Store search popularity by country and genre, impression share and rank, ASO keyword ideas, or deciding which App Store keywords to add, pause, negate, or bid on. Read-only; it never changes campaigns.