ChatGPT Browser Research — Skill
When to Use
- Deep research queries requiring ChatGPT's reasoning
- Getting documentary/narrative structure for content projects
- Second-opinion research alongside Gemini or Claude
- When user explicitly requests ChatGPT for a task
Prerequisites
- Chrome MCP tools available (Claude in Chrome extension)
- User logged into ChatGPT account
- Tab group context established via
tabs_context_mcp
Phase 1: Setup
1.1 Get or Create Tab
tabs_context_mcp(createIfEmpty: true)
If no ChatGPT tab exists:
tabs_create_mcp()
Note the tabId.
1.2 Navigate to ChatGPT
navigate(url: "https://chatgpt.com", tabId: <tabId>)
Wait 3-5 seconds for full load.
1.3 Read Interactive Elements
read_page(tabId: <tabId>, filter: "interactive")
- Identify the prompt textbox (usually a
textbox element)
- Note: ChatGPT's input IS a standard textbox —
form_input may work, but computer(type) is more reliable
Phase 2: Send Research Query
computer(action: "left_click", ref: <textbox_ref>, tabId: <tabId>)
2.2 Type the Query
computer(action: "type", text: "<research_query>", tabId: <tabId>)
Query Crafting Tips for Research:
- Be specific about output format: "structured narrative", "timeline", "comparison table"
- Mention the end use: "for a visual documentary", "for comic panel scripting"
- Ask for cinematic/visual details if generating content for image prompts
- Request "lesser-known details" or "dramatic moments" for richer material
2.3 Submit
computer(action: "key", text: "Return", tabId: <tabId>)
3.1 Wait for Response
computer(action: "wait", duration: 20-30, tabId: <tabId>)
- ChatGPT responses can take 15-45 seconds for long research queries
- Check the tab title — it often updates to reflect the conversation topic when done
3.2 Verify Completion
computer(action: "screenshot", tabId: <tabId>)
- Look for the input box reappearing at the bottom (means response is complete)
- If still generating (animated dots visible), wait another 10 seconds
get_page_text(tabId: <tabId>)
- Returns the full conversation text including the response
- This is the primary extraction method — captures everything
3.4 Save Research Output
Save extracted text to project files for use in content pipelines:
Write tool → save to research/ or docs/ directory
Phase 4: Follow-Up Queries
4.1 Ask for More Detail
Same conversation persists — just type follow-up in the same input:
computer(action: "left_click", ref: <textbox_ref>, tabId: <tabId>)
computer(action: "type", text: "<follow_up_query>", tabId: <tabId>)
computer(action: "key", text: "Return", tabId: <tabId>)
4.2 Useful Follow-Ups for Content Projects
- "Give me 5 cinematic scenes that would visually carry this"
- "Break this into panel-by-panel image prompts in sepia documentary style"
- "What are the most visually dramatic moments in this story?"
- "Create a timeline with exact dates I can use for captions"
Gotchas & Failure Modes
| Issue | Solution |
|---|
| Page takes long to load | Wait 5s after navigate, then read_page |
| Response still generating | Screenshot to check, wait if animated dots visible |
| Refs become stale | Re-read page with read_page after any navigation |
| get_page_text returns too much | Filter the output — ChatGPT page includes sidebar text |
| Model selector defaults to "Auto" | Usually fine; if specific model needed, click model selector first |
| Rate limiting | ChatGPT may throttle — wait and retry |
| Login required | User must be pre-logged in; cannot enter credentials |
Parallel Research Pattern
For maximum research speed, run ChatGPT and Gemini simultaneously:
- Create two tabs (one per AI)
- Navigate both in parallel
- Send queries to both (can use different angles on same topic)
- Wait for both to complete
- Extract text from both with
get_page_text
- Synthesize findings from both sources
This gives you two independent research perspectives to cross-reference.
Parallel & Multi-Session Work
If generating images through ChatGPT while another session uses Gemini, see skills/multi-session-coordination.md for browser lock coordination. Only one session touches the browser at a time.
Cleanup After Download
If downloading any files (images, documents) from the browser:
- Move immediately from Downloads to the project folder
- Delete the original from Downloads
- Downloads is transit, not storage.