Browser Automation Skill
Setup
pip install playwright && playwright install chromium
pip install playwright-stealth
Skill Comparison
| Skill | Use Case |
|---|
browser-automation | Scripted automation, data pipelines, form bots |
webapp-testing | QA assertions, visual regression, test suites |
chrome-browser | Interactive MCP DevTools control |
Iron Law
NEVER USE time.sleep() — use page.wait_for_selector() or wait_for_load_state()
Core Patterns
Navigation and Extraction
with sync_playwright() as p:
context = p.chromium.launch(headless=True).new_context(
user_agent="Mozilla/5.0 ...", viewport={"width": 1280, "height": 720})
page = context.new_page()
page.goto(url); page.wait_for_load_state("networkidle")
items = [i.inner_text() for i in page.get_by_role("listitem").all()]
Form Filling
page.get_by_label("Email").fill("user@example.com")
page.get_by_label("Password").fill("secret")
with page.expect_navigation():
page.get_by_role("button", name="Submit").click()
Screenshot and PDF
page.screenshot(path="full.png", full_page=True)
page.pdf(path="page.pdf", format="A4", print_background=True)
Auth Session Management
context.storage_state(path="auth-state.json")
context = browser.new_context(storage_state="auth-state.json")
Anti-Detection
browser = p.chromium.launch(args=["--disable-blink-features=AutomationControlled"])
context = browser.new_context(user_agent="...", locale="en-US", timezone_id="America/New_York")
page.add_init_script("Object.defineProperty(navigator, 'webdriver', {get: () => undefined})")
Agent Usage
- Save extracted data to
.claude/context/tmp/ as structured JSON
- Save screenshots/PDFs to
.claude/context/artifacts/
- Never commit
auth-state.json — add to .gitignore
Anti-Patterns
| Anti-Pattern | Correct Approach |
|---|
time.sleep() for waits | page.wait_for_selector() / wait_for_load_state() |
| CSS selectors for forms | get_by_label(), get_by_role() |
| Re-login each run | context.storage_state() persistence |
| Committing auth state | Add auth-state.json to .gitignore |
Memory Protocol (MANDATORY)
Before starting: Read .claude/context/memory/learnings.md
After completing:
- New extraction pattern ->
.claude/context/memory/learnings.md
- Bot detection encountered ->
.claude/context/memory/issues.md
ASSUME INTERRUPTION: If it's not in memory, it didn't happen.
