browser-automation – Agents & Automation skill for AI agents
Build reliable browser checks using observed UI state, semantic locators, bounded waits, isolated test data and explicit outcome verification.
Install the browser-automation skill with one command
Pick the tool you use, copy the command and paste it in your terminal. The skill lands in that tool's skills folder and is ready in your next session.
Claude Code
mkdir -p ~/.claude/skills && curl -fsSL https://skill.dn1nguoi.com/z/browser-automation-8337a.tgz | tar xz -C ~/.claude/skills
Call it in Claude Code: /browser-automation – then describe the task you want done.
Codex
mkdir -p ~/.codex/skills && curl -fsSL https://skill.dn1nguoi.com/z/browser-automation-8337a.tgz | tar xz -C ~/.codex/skills
Call it in Codex: $browser-automation – then describe the task you want done.
Hermes Agent
mkdir -p ~/.hermes/skills && curl -fsSL https://skill.dn1nguoi.com/z/browser-automation-8337a.tgz | tar xz -C ~/.hermes/skills
Call it in Hermes Agent: /browser-automation – then describe the task you want done.
DeepSeek
mkdir -p ~/.deepcode/skills && curl -fsSL https://skill.dn1nguoi.com/z/browser-automation-8337a.tgz | tar xz -C ~/.deepcode/skills
Call it in DeepSeek: /browser-automation – then describe the task you want done.
SKILL.md content
Browser Automation
Use the browser tool already selected by the user or installed in the project. Playwright, Puppeteer and Selenium have different integrations; choose from actual requirements rather than unsupported success-rate claims. Modified by AAS maintainers on 2026-09-05: removed unverified comparisons and bypass defaults, clarified waiting and evidence limits.
Separate tests of applications you control from interaction with an existing authenticated browser. Do not replace the latter with a fresh unauthenticated session or extract credentials to make an automation test work.
Detailed Guide
Read [the detailed guide](references/detailed-guide.md) before executing this skill. It retains the complete procedure and reference material. Treat its safety, prerequisites, and validation requirements as mandatory. For focused work, load the relevant sections; for end-to-end work, read the guide completely.
Playwright Test Example
"""
import { test, expect } from '@playwright/test';
// Each test runs in isolated browser context
test('user can add item to cart', async ({ page }) => {
// Fresh context - no cookies, no storage from other tests
await page.goto('/products');
await page.getByRole('button', { name: 'Add to Cart' }).click();
await expect(page.getByTestId('cart-count')).toHaveText('1');
});
test('user can remove item from cart', async ({ page }) => {
// Completely isolated - cart is empty
await page.goto('/cart');
await expect(page.getByText('Your cart is empty')).toBeVisible();
});
"""
Good Examples (User-Facing)
"""
// By role - THE BEST CHOICE
await page.getByRole('button', { name: 'Submit' }).click();
await page.getByRole('link', { name: 'Sign up' }).click();
await expect(page.getByRole('heading', { name: 'Dashboard' })).toBeVisible();
await page.getByRole('textbox', { name: 'Search' }).fill('query');
// By text content
await expect(page.getByText('Welcome back')).toBeVisible();
await page.getByText(/Order #\d+/).click(); // Regex supported
// By label (forms)
await page.getByLabel('Email address').fill('[email protected]');
await page.getByLabel('Password').fill('secret');
// By placeholder
await page.getByPlaceholder('Search...').fill('query');
// By test ID (when no user-facing option works)
await page.getByTestId('submit-button').click();
"""
Bad Examples (Fragile)
"""
// DON'T - CSS selectors tied to structure
await page.locator('.btn-primary.submit-form').click();
await page.locator('#header > div > button:nth-child(2)').click();
// DON'T - XPath tied to structure
await page.locator('//div[@class="form"]/button[1]').click();
// DON'T - Auto-generated selectors
await page.locator('[data-v-12345]').click();
"""
When to Use
Use to verify a real browser workflow, diagnose a UI timing failure or collect explicitly authorized page data. Inspect the current page and available tool APIs before selecting locators or actions.
Worked example and prerequisites
Input: exporting a reviewed JSON file from a local web app. Have the app running, the intended browser available and a synthetic form value. Observe the export control, register the download event before clicking, inspect the downloaded JSON and confirm that editing the input invalidates the old preview. Expected: one file containing the reviewed value, with no hidden project data or network submission.
Use explicit desktop/mobile viewports and inspect keyboard/focus behavior when the task includes usability. A locked desktop leaves interactive verification pending; unit tests and headless probes are separate evidence.
Limitations
- Auto-waiting checks actionability, not business correctness or successful backend writes.
- Screenshots, HTML, traces and auth-state files may contain private information; capture only the authorized scope.
- Retrying a read can be safe; retrying checkout, deletion or sending a message may duplicate a side effect. Verify state before repeating it.
- Resource blocking and mocked responses change the environment and cannot establish unmodified production behavior.
- Examples require the project’s imports, runner and fixture routes; no browser, service or account is installed by this skill.
Related skills
- browser-bridge – Inspect, control, and automate tasks in your real Microsoft Edge or Google Chrome…
- browser-extension-builder – Expert in building browser extensions that solve real problems - Chrome, Firefox, and…
- pi-agent – Builds with and operates Pi, the minimal terminal coding harness. Use for installing Pi…
- browser-extension-reverse – Authorized reverse engineering of Chrome/Firefox extensions: manifest analysis…
- browser-harness – Drive an existing browser through CDP for authenticated, visual, or interactive web…
- browser-testing-with-devtools – Test browser apps with Chrome DevTools MCP by inspecting live DOM, console logs, network…
- bug-bounty – Complete bug bounty workflow
- bug-hunt-swarm – Parallel read-only multi-agent root-cause investigation for bugs, regressions, crashes…