Log in and navigate
Reach a page that sits behind authentication.“Go to app.example.com, log in with my credentials, and open the Reports page.”The agent navigates to the login screen, snapshots it to find the form fields, enters your credentials, submits, waits for the page to load, and then moves on to the target page. If you’re already signed in from an earlier session, it reuses that session and skips the login entirely.
Fill out a complex form
Complete a multi-field form with mixed input types.“Fill out the support ticket form on our internal tool — set priority to High, category to Billing, and describe the issue as ‘Customer charged twice for subscription’.”The agent snapshots the form, matches fields by their labels, types into the text inputs, selects the right dropdown values, and submits.
Extract structured data
Pull information off a page into a usable format.“Open our analytics dashboard and extract the top 10 metrics from the summary table.”Here the agent runs JavaScript in the page context to query the DOM directly, then formats the result as a table or an exportable file. For larger or repetitive extractions, it may switch to calling the page’s backend — see API discovery.
Inspect network traffic
See which API calls a page makes when you interact with it.“Open the dashboard, click Refresh, and show me what API calls it makes.”The agent monitors network activity during the interaction and reports the request URLs, HTTP methods, status codes, and response sizes. This is a great first step before asking it to call those endpoints directly.
Capture a targeted screenshot
Grab a specific element instead of the whole page.“Take a screenshot of just the revenue chart on the dashboard.”The agent locates the element — by reference, CSS selector, or coordinates — and captures only that region.
Trigger keyboard shortcuts
Drive power-user flows through the keyboard.“Open the command palette with Cmd+K and search for ‘billing’.”The agent sends the shortcut, including modifier keys like Shift, Control, Alt, and Meta/Cmd, then continues typing or navigating.
Chain a multi-step workflow
Combine several actions into one task.“Go to our HR portal, open each team member’s profile, and compile their job titles and departments into a spreadsheet.”The agent works through the sequence, collecting data as it goes and assembling the final spreadsheet. On a task this repetitive, it will often notice the pattern and switch to the app’s internal API to finish far faster.
Writing good browser prompts
Be specific about the output
Be specific about the output
Say whether you want a spreadsheet, a table, a file, or just a summary. The agent formats results to match.
Mention your auth state
Mention your auth state
Tell the agent if you’re already signed in, or if it should expect a login wall, so it doesn’t misread the page.
Describe unusual page structure
Describe unusual page structure
If a form uses custom widgets, infinite scroll, or nested tabs, a one-line hint helps the agent target the right elements.
Name the exact fields and values
Name the exact fields and values
For form filling, spell out each field and the value it should hold to avoid ambiguity.
All of these run through the desktop browser. See Browser Automation for requirements and window behavior, and Permissions for how state-changing actions are approved.