Browser Automation

Use the browser capability already available in the environment when possible. To operate or debug a live page from the shell, prefer agent-browser; for native desktop apps, use agent-desktop. This skill covers Playwright code, extraction design, and Manifest V3 extensions. Inspect before acting, preserve the user's session, and distinguish read-only navigation from consequential actions such as submitting, publishing, purchasing, deleting, or sending.

Plan the Interaction

1. Confirm the target, desired artifact, account context, and completion condition.
2. Check site terms, robots policy where relevant, rate limits, and authorization.
3. Classify actions:
   • read-only: open, search, inspect, extract;
   • reversible state change: drafts, filters, saved preferences;
   • consequential: submit, send, purchase, publish, delete, change access.
4. Ask before consequential actions unless the user's request already clearly authorizes that exact action and target.
5. Define a recovery path for any mutation.

Do not bypass CAPTCHAs, access controls, paywalls, bot protections, or account restrictions. Do not reuse credentials outside their intended origin.

Reliable Automation