browser-automation

星标11

分支3

更新时间2026年7月4日 08:43

Use when a task needs a real browser — navigating sites, filling and submitting forms, clicking through flows, extracting page data, or taking screenshots — driving whatever browser tool or MCP server is available.

安装

用 Codex 或 Claude 帮你安装复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。

来源

imzodev

imzodev/openaidy

打开 GitHub 仓库查看创作者相关仓库

下载

Browser Automation

Some things only exist behind a page: a login-walled dashboard, a form with no API, a site that renders with JavaScript. Drive a browser to do them — deliberately, one observable step at a time.

Before you start

Confirm a browser tool or configured MCP server (Playwright, Puppeteer, a browser MCP, etc.) is available. If not, tell the user what to set up instead of pretending to click.
Restate the goal as a concrete end state — "reach the order-confirmation page", "extract the table rows". You need to know when you're done.
Flag any step that spends money, sends a message, or changes someone else's data, and get explicit approval before running it.

Method

Navigate, then observe before acting. Read the actual page — visible text, form fields, current URL — rather than assuming the layout.
Prefer stable selectors: visible labels, roles, and text over brittle CSS/xpath tied to generated class names.
Do one action, then re-check the result. Don't fire a long sequence blind and hope it worked.
Wait for the page to settle — network idle, element present — instead of fixed-duration sleeps.
Capture a screenshot at decision points and on failure. It's your evidence of what actually happened.

Guardrails

Never submit a purchase, booking, message, or account change without confirming first.
Don't enter credentials the user didn't give you; use the browser's existing session or profile where possible.
Stop and report if you hit a login wall, a CAPTCHA, or a page you didn't expect — don't thrash.

See references/automation-recipe.md.

Anti-patterns

Assuming a selector still exists instead of checking the current page.
Chaining many clicks with no verification between them.
Treating a fixed "wait 5 seconds" as a substitute for a real condition.
Auto-accepting consent or purchase dialogs that have real consequences.

同仓库更多 Skills

同仓库

addon-ui-components-cookbook

imzodev/openaidy

Use when writing or reviewing OpenAidy addon UI code (app/index.html, app/index.js) that calls sdk.ui.* — covers component mounting semantics and composed usage patterns beyond the one-line reference in addon_create's tool description.

2026-07-2011

data-and-apis

imzodev/openaidy

Use when calling external REST APIs or querying databases on the user's behalf through available tools or MCP servers — read-only by default, parametrized queries, and confirmation before any write.

2026-07-0411

working-with-documents

imzodev/openaidy

Use when reading, extracting from, or generating real-world document files — PDFs, spreadsheets (CSV/Excel/Sheets), and slide decks — with available tools; preserve the source and verify the output.

2026-07-0411

scheduling

imzodev/openaidy

Use when working with calendars — checking availability, finding meeting times, and creating or updating events through an available calendar integration — with careful timezone handling and confirmation before inviting others.

2026-07-0411

sending-messages

imzodev/openaidy

Use when posting to chat or email on the user's behalf — Slack, Discord, Telegram, or mail — through an available integration. Get the recipient and content right, and confirm before anything goes out.

2026-07-0411

web-research

imzodev/openaidy

Use when you need current or external facts — searching the web, reading multiple sources, cross-checking claims, and synthesizing an answer with citations instead of guessing from memory.

2026-07-0411

name	Browser Automation
description	Use when a task needs a real browser — navigating sites, filling and submitting forms, clicking through flows, extracting page data, or taking screenshots — driving whatever browser tool or MCP server is available.
version	1.0.0

Browser Automation

Some things only exist behind a page: a login-walled dashboard, a form with no API, a site that renders with JavaScript. Drive a browser to do them — deliberately, one observable step at a time.

Before you start

Confirm a browser tool or configured MCP server (Playwright, Puppeteer, a browser MCP, etc.) is available. If not, tell the user what to set up instead of pretending to click.
Restate the goal as a concrete end state — "reach the order-confirmation page", "extract the table rows". You need to know when you're done.
Flag any step that spends money, sends a message, or changes someone else's data, and get explicit approval before running it.

Method

Navigate, then observe before acting. Read the actual page — visible text, form fields, current URL — rather than assuming the layout.
Prefer stable selectors: visible labels, roles, and text over brittle CSS/xpath tied to generated class names.
Do one action, then re-check the result. Don't fire a long sequence blind and hope it worked.
Wait for the page to settle — network idle, element present — instead of fixed-duration sleeps.
Capture a screenshot at decision points and on failure. It's your evidence of what actually happened.

Guardrails

Never submit a purchase, booking, message, or account change without confirming first.
Don't enter credentials the user didn't give you; use the browser's existing session or profile where possible.
Stop and report if you hit a login wall, a CAPTCHA, or a page you didn't expect — don't thrash.

See references/automation-recipe.md.

Anti-patterns

Assuming a selector still exists instead of checking the current page.
Chaining many clicks with no verification between them.
Treating a fixed "wait 5 seconds" as a substitute for a real condition.
Auto-accepting consent or purchase dialogs that have real consequences.