Agent-Skills.md

Agent Skills: use-web-skill

Discover and use browser-advertised skills exposed through HTML head tags like `<link rel="web-skill" href=".../SKILL.md" title="...">`. Use when Codex is browsing or automating a website with Browser MCP, Playwright, or other browser tooling and should check whether the page publishes a task-relevant skill before continuing with normal interaction.

UncategorizedID: dimbreak/web-skill/use-web-skill

Repository

dimbreak/web-skill
dimbreakLicense: MIT
1

Install this agent skill to your local

pnpm dlx add-skill https://github.com/dimbreak/web-skill/tree/HEAD/skills/use-web-skill

Skill Files

Browse the full folder contents for use-web-skill.

Download Skill

Loading file tree…

skills/use-web-skill/SKILL.md

Skill Metadata

Name
use-web-skill
Description
Discover browser-advertised `web-skill` links in a page head, fetch the relevant `SKILL.md`, verify the matching runtime entrypoint under `window._web_skills`, and prefer direct task-level calls before falling back to raw browser automation.

use-web-skill

Use this skill when an agent is working on a live webpage with Browser MCP, Playwright, or similar browser tooling.

It helps the agent look for <link rel="web-skill" ...> in the document <head>, read the linked SKILL.md, verify the matching browser runtime entrypoint, and prefer the page's advertised workflow before falling back to generic DOM automation.

This is most useful on sites that already publish task-level browser skills and want agents to discover them reliably.

Workflow

  1. Read the page head and look for:
    • <link rel="web-skill" ...>
    • especially href and title
  2. Compare each link title against the current task.
  3. If a title looks relevant, fetch the linked SKILL.md.
  4. Read the downloaded skill, identify the documented console entrypoint, and check whether the page exposes it under window._web_skills.
  5. If the matching runtime entrypoint exists, prefer direct task-level function calls through window._web_skills.* instead of recreating the flow with DOM clicks and typed input.
  6. If no relevant web-skill link exists, or the linked skill cannot be used, continue with the normal browser workflow.

Selection rules

  • Prefer the most task-specific title.
  • Prefer exact domain or product matches over generic titles.
  • If multiple links look useful, start with the closest match rather than loading everything.
  • Ignore links with vague titles unless the task is also vague.

Fetch rules

  • Resolve relative href values against the current page URL.
  • Treat the linked file as the source of truth for browser-specific workflow.
  • If the file cannot be fetched, note that and fall back to standard browsing.

Runtime checks

  • After reading the linked SKILL.md, verify that the described runtime key exists under window._web_skills.
  • Sanity-check that the link title, skill description, and runtime entrypoint all point to the same workflow before invoking anything.
  • If the skill document is present but the runtime entrypoint is missing, treat that as an incomplete integration and fall back carefully.

Browser MCP guidance

  • Use Browser MCP or the active browser tool to inspect <head> content first.
  • Only switch into the linked skill workflow after confirming the title is relevant and the runtime entrypoint is available.
  • Prefer direct window._web_skills.* calls whenever the page exposes them.
  • Keep the fallback order simple:
    1. discover <link rel="web-skill">
    2. fetch the linked SKILL.md
    3. verify the runtime entrypoint under window._web_skills
    4. direct-call the task-level function
    5. only then fall back to raw DOM automation

Output expectation

After discovering a relevant web-skill, continue the task using that downloaded SKILL.md as the active operating guide.