Ask Gemini (Google)

Ask Google Gemini a question via browser automation at gemini.google.com.

Prerequisites

• User must already be logged into Google account in browser (check for account avatar in top-right)
• Browser session already initialized (browser_navigate required)

Steps

1. Navigate to Gemini:

   browser_navigate(url='https://gemini.google.com')
   

2. Verify the correct page loaded: Immediately check browser_snapshot. If the page is NOT Gemini (e.g., landed on AssetSentry, Facebook, etc.), the browser has a tab-switching issue:

   # Check all open tabs via CDP
   browser_cdp(method='Target.getTargets', params={})
   # Navigate again with full URL including /app if needed
   browser_navigate(url='https://gemini.google.com/app')
   

3. Wait for page to load and find input field. Use browser_snapshot to locate the textarea. Look for a textbox labeled "Enter a prompt for Gemini". Note its ref (e.g., @e17). Also note the send button ref (e.g., @e13).

4. Type the question:

   browser_type(ref='@<input_ref>', text='<your question>')
   

5. Submit: Click the send button OR press Enter:

   browser_click(ref='@<send_button_ref>')
   # OR
   browser_press(key='Enter')
   

6. Wait for response: Gemini's response may take multiple polling cycles before rendering. Poll with browser_vision or browser_snapshot until you see the response text appear (indicated by a sparkle/typing indicator disappearing).

7. Read the answer — use browser_console for long responses (RECOMMENDED):

   browser_console(expression='document.querySelector("main")?.innerText || document.querySelector(".wEwyrc")?.innerText || "not found"')
   

   ⚠️ The old selector document.querySelector("main").innerText often returns null and fails. Always use optional chaining (?.) and fallback selectors.

   If browser_console fails or returns empty, fall back to browser_vision:

   browser_vision(question='Read ALL of Gemini\\'s response text from the beginning. Include ALL sections, bullet points, and data. Don\\'t stop until you\\'ve read everything Gemini wrote.')
   

8. Send text result via Discord (default — no screenshot):

   send_message(message='<formatted text summary>', target='discord')
   

   除非使用者特別要求截圖，否則預設只傳文字結果。

Important Notes

• Always verify page URL after navigation. The browser tool may land on a different tab (AssetSentry, Facebook, etc.) — check browser_snapshot title/URL immediately.
• CDP for tab debugging: If navigation lands on the wrong page, use browser_cdp(method='Target.getTargets') to inspect all open tabs.
• browser_console requires robust selectors: Never use document.querySelector("main").innerText (it returns null). Use ?. optional chaining and fallback selectors.
• 預設不傳截圖，只傳送文字結果摘要。使用者明確要求時才額外傳截圖。
• If the send button is disabled, the input might be empty. Try clicking elsewhere first.
• Gemini's response rendering can be delayed — poll with vision/snapshot until text appears.
• The page may show a "Temporary chat" banner near the top — this is normal, proceed regardless.

When to Use

• User asks to ask Gemini a question
• User wants to interact with Google Gemini
• User wants to pose a question or get information from Gemini
• User asks to chat with or query Gemini

When NOT to Use

• Use web_search for simple factual queries
• Use google-workspace skill for Gmail/Calendar/Docs tasks
• Use browser skill only if user needs general browser navigation tasks