Playwright-Efficient Web Research

Use this skill for web exploration tasks where context window efficiency matters.

Use This Skill When

• The user asks to research web pages, compare content across sites, or extract specific facts.
• Browser interaction is required (click, type, login flow, pagination, dynamic content).
• Repeated full HTML/text fetch would waste tokens.

CLI vs MCP Boundary

• Default: Playwright CLI (playwright-cli) with direct commands (no wrapper scripts).
• Fallback: Playwright MCP only for long-running/state-heavy autonomous loops requiring persistent tool-level introspection.

Core Rules

1. Reuse one named session per task (-s=<session>); avoid reopening browser every step.
2. In this sandbox, pass --browser=chromium explicitly unless you intentionally need another channel.
3. Use snapshot to get element refs, then interact by ref (click e12, fill e8 ...).
4. Extract only required fields with eval; do not dump full page content unless explicitly requested.
5. Capture screenshot/PDF/logs only when evidence is required.
6. Close session when done (close, optional delete-data).
7. Keep sessions non-persistent by default; enable --persistent only when cross-command login state is explicitly required.

Workflow (Direct CLI)

1. Open session once (default: non-persistent):

playwright-cli -s=research-1 open "https://example.com" --browser=chromium

If the workflow explicitly requires saved login/session state:

playwright-cli -s=research-1 open "https://example.com" --browser=chromium --persistent

2. Navigate + interact with refs:

playwright-cli -s=research-1 snapshot
playwright-cli -s=research-1 click e15
playwright-cli -s=research-1 fill e7 "search query"
playwright-cli -s=research-1 press Enter

3. Targeted extraction (minimum needed fields only):

playwright-cli -s=research-1 eval \
  "() => ({ title: document.title, h1: document.querySelector('h1')?.textContent?.trim() ?? null })"
playwright-cli -s=research-1 eval "el => el.textContent?.trim() ?? null" e15

4. Optional evidence:

playwright-cli -s=research-1 screenshot --filename=/tmp/research-1.png
playwright-cli -s=research-1 pdf --filename=/tmp/research-1.pdf
playwright-cli -s=research-1 console
playwright-cli -s=research-1 network

5. Cleanup:

playwright-cli -s=research-1 close
playwright-cli -s=research-1 delete-data

Anti-Patterns

• Repeatedly opening new sessions for each click.
• Fetching complete page text/DOM when only 2-3 fields are needed.
• Using screenshots as default extraction path.
• Adding unnecessary wrapper scripts when direct CLI commands are sufficient.
• Switching to MCP for simple linear navigation tasks.

Notes

• If global playwright-cli is unavailable, run the same commands with npx --yes @playwright/cli ....
• Session names should be task-specific (news-compare, pricing-check, auth-flow-debug).