deep-research

name	deep-research
description	Launch a deep research query across 10 AI platforms simultaneously using browser automation. Opens ChatGPT, Claude, Gemini, Grok, Parallel, Perplexity, Manus, DeepSeek, Mistral, and Kimi in separate browser tabs, enables each platform's deep/expert research mode, enters the user's query, and monitors progress. Use this skill whenever the user says "deep research", "research this across platforms", "multi-platform research", "run research everywhere", "parallel research", "compare AI research", or any variation of wanting to query multiple AI chatbots at once for in-depth research. Also triggers when the user mentions wanting to use multiple AI tools for research, or asks to "search all AIs", "ask all the AIs", or "get research from every platform".
compatibility	[{"tool":"mcp__claude-in-chrome__tabs_context_mcp"},{"tool":"mcp__claude-in-chrome__tabs_create_mcp"},{"tool":"mcp__claude-in-chrome__navigate"},{"tool":"mcp__claude-in-chrome__computer"},{"tool":"mcp__claude-in-chrome__find"},{"tool":"mcp__claude-in-chrome__read_page"},{"tool":"mcp__claude-in-chrome__form_input"},{"tool":"mcp__claude-in-chrome__get_page_text"}]

Deep Research Across 10 AI Platforms

This skill automates launching a deep research query across 10 AI platforms in parallel using Claude in Chrome browser automation. The user provides a research query, and you open all platforms in separate tabs, enable their respective deep research modes, enter the query, and monitor progress.

Prerequisites

The user must already be logged into all 10 platforms in their Chrome browser. Do not attempt to log in on their behalf.

Platform Registry

Each platform has a URL, and a specific sequence of UI actions to enable deep research mode before entering the query. Here they are:

#	Platform	URL	Deep Research Activation
1	ChatGPT	https://chatgpt.com/	Click the "+" button, then select "Deep Research"
2	Claude	https://claude.ai/new	Click the "+" button, then select "Research"
3	Gemini	https://gemini.google.com/app	Click "Tools", then select "Deep Research"
4	Grok	https://grok.com/	Click "Expert" to the right of the search bar
5	Parallel	https://platform.parallel.ai/play/deep-research	Select "Ultra8x" from the dropdown in the search bar
6	Perplexity	https://www.perplexity.ai/	Click the "+" button, then select "Deep Research"
7	Manus	https://manus.im/app	On the bottom right of the search bar, click "More", then select "Wide Research"
8	DeepSeek	https://chat.deepseek.com/	Click "DeepThink" and "Search" under the search bar
9	Mistral	https://chat.mistral.ai/chat	Click "Research" under the search bar
10	Kimi	https://www.kimi.com/deep-research	Verify that "Agent|Deep Research" is already selected

Workflow

Phase 1: Get the Research Query

Ask the user for their research query if they haven't already provided one.

CRITICAL — Query Preservation Rule: Use the user's query EXACTLY as provided. Do NOT rephrase, reword, shorten, expand, or "improve" it in any way. Copy the text character-for-character into every platform. If you feel the query is ambiguous, ask the user to clarify — do not silently modify it.

Phase 2: Open All Tabs

1. Call tabs_context_mcp to get the current browser state.
2. Create 10 new tabs using tabs_create_mcp (one per platform). Record each tab ID.
3. Navigate each tab to its respective platform URL using navigate. Do this for all 10 tabs — you can issue multiple navigate calls in the same turn since they are independent.
4. Wait a few seconds for pages to load (computer with action: wait).

Phase 3: Enable Deep Research Mode (Per Platform)

Work through each tab to enable the platform's deep research mode. Because each platform has a different UI, you need to handle them individually. Take a screenshot of each tab before interacting to understand the current state.

The general approach for each platform:

1. Take a screenshot to see the current UI state.
2. Use find or read_page to locate the relevant UI elements (buttons, dropdowns, menus).
3. Click the activation elements in sequence per the Platform Registry above.
4. Verify the mode is active by taking another screenshot or reading the page.

Tips for specific platforms:

• ChatGPT: The "+" button is typically near the message input area. After clicking it, a menu appears — look for "Deep Research" in the menu items.
• Claude: Similar to ChatGPT — "+" button near the input, then select "Research" from the menu. If the find tool doesn't work to locate the "Research" menuitem, use read_page with filter: "interactive" to get all interactive elements and find the one labeled "Research". HARD GATE — Never fall back to standard chat: If after trying both find and read_page you still cannot locate the "Research" mode option, do NOT submit the query in regular chat mode. Instead, stop, tell the user: "I cannot find the Research mode option on Claude. Please enable it manually (it may need to be turned on in your Claude settings or plan), then let me know when it's ready." Wait for the user to confirm before proceeding with Claude's tab. IMPORTANT: After submitting the research query, Claude will ask for tool use permission (e.g., "Allow Claude to use web search?" or similar). The user must manually accept this in their browser — remind the user to allow it when you see it, or note in your status report that Claude needs tool use approval.
• Gemini: "Tools" is typically a button/tab in the UI. After clicking, look for "Deep Research" option. Gemini may show a research plan — accept the default plan without modification.
• Grok: "Expert" appears as a mode selector near/right of the search bar. It may be a toggle or button.
• Parallel: There's a dropdown in the search bar area. Click it and select "Ultra8x".
• Perplexity: "+" button near the search input, then "Deep Research" from the resulting menu.
• Manus: IMPORTANT — The "More" button is NOT an icon on the search bar itself. There is a row of suggestion cards/buttons at the bottom of the chat area (e.g., "Travel Planning", "Data Analysis", "More"). The "More" button is the LAST item in this row of suggestion cards. Do NOT click the puzzle icon (that opens skills/plugins panel) or the model dropdown (that shows plan options). Click the "More" text in the suggestion cards row at the bottom. After clicking "More", a dropdown menu appears with "Wide Research" at the top — click it.
• DeepSeek: "DeepThink" and "Search" are separate toggles/buttons below the search bar. Enable both.
• Mistral: "Research" appears as a mode option under the search bar. Click to select it.
• Kimi: The URL already targets deep research. Just verify "Agent|Deep Research" is selected in the UI. If not, select it.

Process platforms in batches to balance speed with reliability. A good approach: handle 3-4 platforms, then move to the next batch. Take screenshots liberally — browser UIs change and you need to see what you're clicking.

Phase 4: Enter the Query and Submit

For each platform tab:

1. Locate the text input / search bar / message box.
2. Click on it to focus.
3. Type the research query using computer with action: type.
4. Press Enter or click the submit/send button.

Move through all 10 tabs to enter and submit the query. Some platforms use a textarea, others use a contenteditable div — use find to locate the input element if clicking in the expected location doesn't work.

Phase 5: Handle Follow-up Questions

Some platforms ask clarifying questions before starting the research. After submitting queries in all tabs:

1. Cycle through each tab and take a screenshot.
2. If a platform is asking a follow-up question or showing a confirmation dialog:
   • Gemini research plan: Accept the default plan. Click the accept/confirm button.
   • Any other platform asking a question: Report the question to the user in the chat. Tell them which platform is asking. Wait for the user's answer, then go back to that tab and provide the answer.
3. If a platform has already started researching (showing progress indicators, thinking animations, etc.), note it as "running" and move on.

Phase 6: Monitor and Report

After all queries are launched:

1. Wait 30-60 seconds, then cycle through all tabs taking screenshots.
2. Report to the user which platforms are:
   • Still researching (show progress if visible)
   • Finished (have produced results)
   • Stuck or errored (need attention)
3. If any platform needs the user's input (follow-up questions that appeared after launch), surface those.
4. Continue monitoring every 60 seconds until the user tells you to stop, or all platforms have completed.

When reporting, keep it concise:

Research Status Update:
✅ Perplexity — Complete
✅ Gemini — Complete
🔄 ChatGPT — Still researching (~60% through)
🔄 Claude — Still researching
🔄 Grok — Still researching
❌ Manus — Needs your input (asking: "What specific aspects of X should I focus on?")
🔄 DeepSeek — Still researching
🔄 Mistral — Still researching
✅ Kimi — Complete
🔄 Parallel — Still researching

Error Handling

• Tab fails to load: Retry navigation once. If it still fails, report to the user and continue with other platforms.
• Platform not logged in: If you see a login screen, tell the user which platform needs authentication and skip it. Do not attempt to log in.
• UI element not found: Take a screenshot, try find with different descriptions, and try read_page with interactive filter. If the element truly can't be found, report that the UI may have changed.
• Platform timeout: If a platform hasn't responded in 5 minutes during setup, skip it and report.

Important Notes

• Always use find or read_page to locate elements rather than hard-coding coordinates. Platform UIs change frequently and coordinates will break.
• Take screenshots before and after major actions to verify success.
• When typing the query, be aware that some platforms may have character limits or may interpret certain characters specially. Type the query exactly as provided by the user — character for character, including any unusual capitalisation, line breaks, bullet style, or apparent typos. Do not normalise or improve the text.
• If a platform shows a CAPTCHA or bot detection, inform the user immediately — they'll need to handle it manually.

Lessons from Live Testing

These tips come from actual test runs and address common pitfalls:

1. Manus "More" button misidentification: The Manus UI has multiple icons that look like they could be "More" — a puzzle icon (opens skills panel) and a model dropdown. The actual "More" is a text label in the suggestion cards row at the BOTTOM of the chat area, alongside cards like "Travel Planning" and "Data Analysis". Always scroll down and look for the suggestion cards row.

2. Claude tool use permission: After submitting a research query on Claude, it will prompt the user to allow tool use (web search, etc.). This is a browser-level permission dialog that the user must click "Allow" on manually. Remind the user to do this in your first status report, or it will sit waiting.

3. Claude Research mode menu: The "+" dropdown menu on Claude can be finicky — clicking by coordinates may not work reliably as the menu closes quickly. Use the find tool to locate the menuitemcheckbox for "Research" and click it via ref instead of coordinates.

4. Parallel detach on long text: Typing very long queries on Parallel may cause the browser tab to detach. If type fails, use javascript_tool to set the value directly on the contenteditable element: el.innerText = "...query..."; el.dispatchEvent(new Event('input', { bubbles: true }));

5. Parallel START RESEARCH button: After entering the query, Parallel requires clicking the orange "START RESEARCH →" button — pressing Enter alone does not start the research.

6. Kimi follow-up questions: Kimi frequently asks 5 clarifying questions before starting deep research. Surface these to the user or answer with reasonable defaults (open to all themes, any project size, open to collaboration, both Spanish and English, etc.) and instruct Kimi to proceed.

7. Gemini auto-accept: Gemini shows a research plan with "Edit plan" and "Start research" buttons. Per the skill instructions, always click "Start research" to accept the default plan without modification.

8. Perplexity input element: Perplexity uses a contenteditable <div> not a <textarea>. If type doesn't work, use javascript_tool to set innerText on the [contenteditable="true"] or [role="textbox"] element.

9. Mistral "Start research" button: After entering the query, Mistral's Research mode generates a research plan but does NOT auto-start execution. It displays "Edit" and "Start research" buttons. You MUST scroll down to find and click the "Start research" button explicitly, or Mistral will sit idle with the plan visible but no research running.

10. Kimi quota exhaustion: Kimi Deep Research has a usage quota that may be exhausted. When this happens, it shows "Quota used up" and "Starting Deep Research creates a new chat". There is no workaround — create a placeholder file with the conversation link and note that quota was exhausted.

11. Chrome extension disconnects during long waits: The Claude in Chrome extension frequently disconnects during 30-second wait actions. After any wait of 15+ seconds, always call tabs_context_mcp to reconnect before issuing further browser commands. If you get connection errors, tabs_context_mcp will re-establish the connection.

12. Sibling tool call failures: When making multiple parallel browser tool calls (e.g., get_page_text on 3 tabs simultaneously), if ONE call fails, ALL sibling calls in that batch also fail. Always retry failed calls individually rather than in another batch.

13. ChatGPT canvas not extractable: ChatGPT Deep Research produces results in a "canvas" document panel that is NOT accessible to browser automation. get_page_text returns only the user's query text. read_page shows internal links, not actual content. javascript_tool is blocked on chatgpt.com. Clicking expand, download, or title icons does not help. Workaround: Create a placeholder file with the ChatGPT conversation link and instruct the user to copy the content manually.

14. Manus document extraction: get_page_text fails on Manus with "No semantic content element found and page body is too large". The fix: at the bottom of the Manus chat, there is a document thumbnail/card showing the final report title (e.g., "Produce and deliver the final comprehensive report"). Click this card — it opens a document viewer modal. Then use read_page with depth: 15 and max_chars: 80000+ on the modal to extract the full report content from the accessibility tree.

15. Query must be used verbatim: Never paraphrase or rewrite the user's research query. Enter it character-for-character on every platform. Editing queries — even with good intentions — changes the research results and violates user intent.

16. Claude Research mode is mandatory: If the Research mode menu item cannot be found on Claude after exhausting find and read_page approaches, do NOT fall back to standard chat. Pause and ask the user to enable it manually, waiting for their confirmation before continuing.

Phase 7: Result Extraction and Saving

After all platforms have completed (or timed out / errored), extract results and save them as individual markdown files. Each platform needs a different extraction strategy:

Extraction Strategy Per Platform

Platform	Method	Notes
ChatGPT	❌ Not extractable	Canvas content blocked. Save placeholder with link.
Claude	read_page on artifact panel	Click the artifact button to open the side panel, then read_page with depth: 15 and max_chars: 100000. Content is in the region "Artifact panel: ..." subtree.
Gemini	read_page accessibility tree	Parse the accessibility tree, filtering for content nodes. May need javascript_tool to extract from JSON-like structures.
Grok	get_page_text	Works directly. Returns full report text.
Parallel	get_page_text	Works directly. Content may be in structured key-value format — reformat into clean markdown.
Perplexity	get_page_text	Works directly. Returns full report with citations.
Manus	read_page after clicking document	Click the document thumbnail at bottom of chat, then read_page with high max_chars on the opened viewer.
DeepSeek	get_page_text	Works directly. Returns full report text.
Mistral	read_page with high max_chars	get_page_text may exceed limits. Use read_page with max_chars: 80000+. Content includes search steps, sources, and analysis.
Kimi	get_page_text or placeholder	If completed, get_page_text works. If quota exhausted, save placeholder.

File Naming Convention

Save files as {##}-{platform}.md in the user-specified output directory:

01-chatgpt.md
02-claude.md
03-gemini.md
04-grok.md
05-parallel.md
06-perplexity.md
07-manus.md
08-deepseek.md
09-mistral.md
10-kimi.md

File Format

Each file should include:

• Platform name and research mode used
• Stats (sources, time, steps) if available
• Direct link to the conversation/report
• Full extracted content formatted as clean markdown
• For placeholders: note explaining why content couldn't be extracted + conversation link

Use Task (Bash subagent) with heredoc to write large files, as direct Write tool may struggle with very long content.