Playwright (Automation + MCP + Scraper)Playwright自动化

When to Use

Use this skill for real browser tasks: JS-rendered pages, multi-step forms, screenshots or PDFs, UI debugging, Playwright test authoring, MCP-driven browser control, and structured extraction from rendered pages.

Prefer it when static fetch is insufficient or when the task depends on browser events, visible DOM state, authentication context, uploads or downloads, or user-facing rendering.

If the user mainly wants the agent to drive a browser with simple actions like navigate, click, fill, screenshot, download, or extract, treat MCP as a first-class path.

Use direct Playwright for scripts and tests. Use MCP when browser tools are already in the loop, the user explicitly wants MCP, or the fastest path is browser actions rather than writing new automation code.

Primary fit is repo-owned browser work: tests, debugging, repros, screenshots, and deterministic automation. Treat rendered-page extraction as a secondary use case, not the default identity.

Architecture

This skill is instruction-only. It does not create local memory, setup folders, or persistent profiles by default.

Load only the smallest reference file needed for the task. Keep auth state temporary unless the repository already standardizes it and the user explicitly wants browser-session reuse.

Quick Start

MCP browser path

CODEBLOCK0

Use this path when the agent already has browser tools available or the user wants browser automation without writing new Playwright code.

Common MCP actions

Typical Playwright MCP tool actions include:

• - browser_navigate for opening a page
• INLINECODE1 and browser_press for interaction
• INLINECODE3 and browser_select_option for forms
• INLINECODE5 and browser_evaluate for inspection and extraction
• INLINECODE7 for uploads
• screenshot, PDF, trace, and download capture through the active browser workflow

Common browser outcomes

Goal	Typical MCP-style action
Open and inspect a site	navigate, wait, inspect, screenshot
Complete a form

navigate, click, fill, select, submit | | Capture evidence | screenshot, PDF, download, trace | | Pull structured page data | navigate, wait for rendered state, extract | | Reproduce a UI bug | headed run, trace, console or network inspection |

Existing test suite

CODEBLOCK1

Bootstrap selectors and flows

CODEBLOCK2

Direct script path

CODEBLOCK3

Quick Reference

Topic	File
Selector strategy and frame handling	INLINECODE8
Failure analysis, traces, logs, and headed runs

debugging.md | | Test architecture, mocks, auth, and assertions | testing.md | | CI defaults, retries, workers, and failure artifacts | ci-cd.md | | Rendered-page extraction, pagination, and respectful throttling | scraping.md |

Approach Selection

Situation	Best path	Why
Static HTML or a simple HTTP response is enough	Use a cheaper fetch path first	Faster, cheaper, less brittle
You need a reliable first draft of selectors or flows

Start with codegen or a headed exploratory run | Faster than guessing selectors from source or stale DOM | | Local app, staging app, or repo-owned E2E suite | Use @playwright/test | Best fit for repeatable tests and assertions | | One-off browser automation, screenshots, downloads, or rendered extraction | Use direct Playwright API | Simple, explicit, and easy to debug in code | | Agent/browser-tool workflow already depends on browser_* tools or the user wants no-code browser control | Use Playwright MCP | Fastest path for navigate-click-fill-screenshot workflows | | CI failures, flake, or environment drift | Start with debugging.md and ci-cd.md | Traces and artifacts matter more than new code |

Core Rules

1. Test user-visible behavior and the real browser boundary

• - Do not spend Playwright on implementation details that unit or API tests can cover more cheaply.
• Use Playwright when success depends on rendered UI, actionability, auth, uploads or downloads, navigation, or browser-only behavior.

2. Make runs isolated before making them clever

• - Keep tests and scripts independent so retries, parallelism, and reruns do not inherit hidden state.
• Extend the repository's existing Playwright harness, config, and fixtures before inventing a parallel testing shape from scratch.
• Do not share mutable accounts, browser state, or server-side data across parallel runs unless the suite was explicitly designed for it.

3. Reconnaissance before action

• - Open, wait, and inspect the rendered state before locking selectors or assertions.
• Use codegen, headed mode, or traces to discover stable locators instead of guessing from source or stale DOM.
• For flaky or CI-only failures, capture a trace before rewriting selectors or waits.

4. Prefer resilient locators and web-first assertions

• - Use role, label, text, alt text, title, or test ID before CSS or XPath.
• Assert the user-visible outcome with Playwright assertions instead of checking only that a click or fill command executed.
• If a locator is ambiguous, disambiguate it. Do not silence strictness with first(), last(), or nth() unless position is the actual behavior under test.

5. Wait on actionability and app state, not arbitrary time

• - Let Playwright's actionability checks work for you before reaching for sleeps or forced actions.
• Prefer expect, URL waits, response waits, and explicit app-ready signals over generic timing guesses.

6. Control what you do not own

• - Mock or isolate third-party services, flaky upstream APIs, analytics noise, and cross-origin dependencies whenever the goal is to verify your app.
• For rendered extraction, prefer documented APIs or plain HTTP paths before driving a full browser.
• Do not make live third-party widgets or upstream integrations the reason your suite flakes unless that exact integration is what the user asked to validate.

7. Keep auth, production access, and persistence explicit

• - Do not persist saved browser state by default.
• Reuse auth state only when the repository already standardizes it or the user explicitly asks for session reuse.
• For destructive, financial, medical, production, or otherwise high-stakes flows, prefer staging or local environments and require explicit user confirmation before continuing.

Playwright Traps

• - Guessing selectors from source or using first(), last(), or nth() to silence ambiguity -> the automation works once and then flakes.
• Starting a new Playwright structure when the repo already has config, fixtures, auth setup, or conventions -> the new flow fights the existing harness and wastes time.
• Testing internal implementation details instead of visible outcomes -> the suite passes while the user path is still broken.
• Sharing one authenticated state across parallel tests that mutate server-side data -> failures become order-dependent and hard to trust.
• Reaching for force: true before understanding overlays, disabled state, or actionability -> the test hides a real bug.
• Waiting on networkidle for chatty SPAs -> analytics, polling, or sockets keep the page "busy" even when the UI is ready.
• Driving a full browser when HTTP or an API would answer the question -> more cost, more flake, less signal.
• Treating third-party widgets and live upstream services as if they were stable parts of your own product -> failures stop being actionable.

External Endpoints

Endpoint	Data Sent	Purpose
User-requested web origins	Browser requests, form input, cookies, uploads, and page interactions required by the task	Automation, testing, screenshots, PDFs, and rendered extraction
INLINECODE28

Package metadata and tarballs during optional installation | Install Playwright or Playwright MCP |

No other data is sent externally.

Security & Privacy

Data that leaves your machine:

• - Requests sent to the websites the user asked to automate.
• Optional package-install traffic to npm when installing Playwright tooling.

Data that stays local:

• - Source code, traces, screenshots, videos, PDFs, and temporary browser state kept in the workspace or system temp directory.

This skill does NOT:

• - Create hidden memory files or local folder systems.
• Recommend browser-fingerprint hacks, challenge-solving services, or rotating exits.
• Persist sessions or credentials by default.
• Make undeclared network requests beyond the target sites involved in the task and optional tool installation.
• Treat high-stakes production flows as safe to automate without explicit user direction.

Trust

By using this skill, browser requests go to the websites you automate and optional package downloads go through npm.
Only install if you trust those services and the sites involved in your workflow.

Related Skills

Install with clawhub install <slug> if user confirms:

• - web - HTTP-first investigation before escalating to a real browser.
• INLINECODE31 - Broader extraction workflows when browser automation is not the main challenge.
• INLINECODE32 - Capture and polish visual artifacts after browser work.
• INLINECODE33 - Find and shortlist target pages before automating them.

Feedback

• - If useful: INLINECODE34
• Stay updated: INLINECODE35