browser

name	browser
description	Browser automation through Penguin's browser-harness-backed tools. Use for web-app testing, visual verification, authenticated browser workflows, dynamic pages, and browser-specific scraping when static HTTP/doc scripting is insufficient.
allowed-tools	["browser_open_tab","browser_page_info","browser_harness_screenshot","browser_click","browser_type","browser_key","browser_fill","browser_wait","browser_js","browser_list_tabs","browser_switch_tab","read_image"]

Browser Skill

Use this skill when a task requires an actual browser: web-app testing, visual verification, authenticated workflows, dynamic/SPA pages, screenshots, coordinate interaction, iframe/shadow-DOM-heavy UIs, downloads/uploads, or browser-specific debugging.

For ordinary documentation lookup and static scraping, prefer scriptable approaches first: official docs, HTTP fetches, search, or page JavaScript extraction. A real browser is most valuable when you need to verify software Penguin built, interact with stateful UI, or inspect dynamic pages that do not expose useful static HTML.

Penguin Tool Workflow

1. Open a new tab with browser_open_tab; do not clobber a user's active tab.
2. Check browser_page_info to confirm URL/title/tab state.
3. Capture browser_harness_screenshot before and after meaningful visible actions.
4. Use read_image for arbitrary local screenshots/artifacts that did not originate from the browser tool.
5. For web-app testing, prefer screenshot -> coordinate click/type/key -> screenshot verification.
6. For docs/data extraction, prefer browser_js or non-browser scripting over repeated manual clicking.
7. Use DOM/CDP-style inspection only when screenshots/coordinates are the wrong tool: hidden nodes, data extraction, virtual DOM state, network ambiguity, or precise selector reads.

Interaction References

Progressively load references from interaction-skills/ only when the mechanic appears:

• screenshots.md - image sizing, devicePixelRatio, and max dimensions.
• tabs.md - tab selection and stale/internal target recovery.
• connection.md - local Chrome/CDP connection troubleshooting.
• dialogs.md - browser dialogs and modal handling.
• dropdowns.md - native/custom dropdown strategy.
• iframes.md and cross-origin-iframes.md - iframe caveats.
• shadow-dom.md - shadow DOM caveats.
• network-requests.md - ambiguous SPA/network completion.
• profile-sync.md - remote/profile workflows when supported.
• downloads.md, uploads.md, drag-and-drop.md, scrolling.md, viewport.md, cookies.md, print-as-pdf.md for specific mechanics.

Do not bulk-load every reference. Pick the smallest reference that matches the current failure mode.

Domain Skills

Browser domain skills live under domain-skills/<hostname>/ inside this browser skill directory. They are opt-in site playbooks, not global prompt content.

When domain skills are enabled, look for host-specific notes before inventing a site-specific approach. If no host skill exists, continue normally and only add a new domain skill when you discover a durable site behavior worth reusing.

Use the domain skill template and redaction policy in domain-skills/README.md. Domain skills must describe stable public mechanics, not private task diaries.

Safety Rules

• A real browser may be logged into real accounts. Treat it as user state, not a disposable sandbox.
• Do not enter credentials, payment details, secrets, or 2FA codes from screenshots or memory. Ask the user to perform sensitive entry.
• Confirm before purchases, posting, account changes, destructive admin actions, legal/medical/financial submissions, or sending messages.
• Redact secrets before saving learned domain skills. Store selectors, URL patterns, public API shapes, waits, and stable quirks; never store cookies, tokens, account IDs, private URLs with sensitive query params, personal data, or screenshot-derived secrets.