On 15 October 2026, OpenAI turned its Chrome extension into something closer to a real browser agent. ChatGPT can now read across your open tabs in Chat mode, and act inside them in Work mode — filling forms, sending messages, downloading files and grouping tabs. The catch is obvious: you’re handing an AI control of pages where you’re usually already signed in. These ten copy-paste prompts split the upgrade into safe read-only prompts and supervised action prompts — and every one comes with a check, so you verify what the agent actually did instead of trusting it.
How to use these prompts
Every prompt below assumes the extension is connected and the @Chrome tag is in play. The setup, as OpenAI documents it: enable Computer Use in the ChatGPT desktop app’s settings, install the official extension from the Chrome Web Store (check the publisher reads OpenAI — copycat “ChatGPT sidebar” extensions exist), then in any chat type @ and select Chrome from the dropdown. Chat mode answers read-only questions across your tabs; Work mode takes actions inside them. One note on plans: OpenAI’s launch thread says the update is “available for all ChatGPT plans,” while PiunikaWeb reports Work mode needs at least ChatGPT Plus. Treat Plus as the safe assumption and check your own account.
The 5 read-only Chat prompts (start here)
Chat mode only reads your tabs — it can’t click, type or submit, which makes these five the safest way to learn the tool. They do the work OpenAI itself demoed at launch, and they’re where every new user should spend the first day.
1. Cross-tab product comparison
Open three product pages and get one comparison table instead of switching tabs for twenty minutes.
Compare the three couches I have open by dimensions, material, delivery time, and return policy. Put it in a table, cite each tab by name, and flag anything you couldn’t find on the page.
What to check: spot-check two numbers against the actual tabs. Prices, dimensions and delivery windows change mid-page, and the agent reads each page as loaded. (Prompt adapted from OpenAI’s launch-thread samples, via 9to5Google’s reproduction of the thread.)
2. Pricing pages vs your requirements
Stop reading pricing pages one by one — hand ChatGPT your requirements and make it do the elimination.
Compare these pricing pages against our requirements: SSO, 20 seats, EU data hosting. Which plan includes everything we need? Tell me what’s missing on each plan, and flag any price that’s only visible after a demo call.
What to check: every “everything you need” verdict against the plan’s own feature table. Vendors hide limits in footnotes; the demo-call flag catches the most common one.
3. Plain-English explainer for anything open
Any dense article, whitepaper or policy page becomes a five-minute read with this one.
Explain the article I have open in plain English. Identify its main argument, its three strongest pieces of evidence, and one claim that reads more like marketing than fact.
What to check: the “main argument” line against your own skim. This is the prompt that most often exposes a salesy piece hiding behind research-speak.
4. Vendor shortlist into a sheet
Turn a research tab pile into a structured shortlist you can actually send to a colleague.
Put the vendors in these tabs into my comparison sheet, with pricing, key features, and source links. Add a column for deal-breakers and leave it blank where the tab doesn’t say.
What to check: every source link should open the actual vendor tab. Empty deal-breaker cells are information — they tell you which vendor is hiding something.
5. Research triage across tabs
When twenty tabs are open and you don’t know where to start, let it rank your own reading list.
Look at the research tabs I have open about [your topic]. Give me a two-sentence summary of each, rank them by usefulness, and tell me which single tab is worth reading in full.
What to check: the top-ranked tab deserves a real skim — rankings are the agent’s judgement, and judgement is the thing to audit first.
The 5 supervised Work prompts (always review)
Work mode takes actions — so the golden rule is the same on every prompt below: tell it to stop before submitting, and you keep the final click. Reporting on the launch could not confirm whether every submission triggers an approval prompt, which is exactly why OpenAI’s own sample prompts include the stop rule. Copy it.
6. Form fill — stop before submit
The flagship Work-mode demo, and the one where the stop rule matters most.
Use the details in this event brief to fill out the registration form, and stop before submitting. Then tell me which fields you filled from the brief and which ones you left blank.
What to check: read every field before you submit — especially dropdowns and date formats, where briefs are vague and the agent guesses. (Prompt from OpenAI’s launch thread, via Unite.AI.)
7. Form + calendar double action
Two actions, one prompt — and a review step between them.
Fill out the form I have open in @chrome ahead of seeing my new dentist. Also, can you add the appointment to my work and personal calendar — but ask me before adding anything.
What to check: confirm the calendar entry has the right date and timezone before you approve it. Timezone is the silent killer of agent-made appointments.
8. Compare, then draft — never send
Work mode can draft outreach from your tab research; “draft” means it stays in the chat, not the outbox.
Look at all the tabs I have open and compare the contractor services. Then draft an email to the one that fits my home renovation — write the draft in this chat, do not send it.
What to check: you press send, not the agent. “Do not send” is belt-and-braces — OpenAI’s launch copy says it “can send messages,” which is precisely why the instruction must be explicit.
9. Tab cleanup with confirmation
The safest-feeling Work task — until it touches a pinned tab. Make confirmation part of the prompt.
Group all my open tabs into topic-based groups. Before you move or close anything, show me the proposed groups and ask for my OK.
What to check: pinned tabs and tab order after it runs. Grouping is a Work capability OpenAI listed at launch (“organize and group that pile of tabs”), but your pinned research tabs are sacred — the confirmation step exists for them.
10. Batch downloads with a manifest
Downloading is a Work capability, and it runs under permission prompts — so ask for the manifest first, downloads second.
Download the invoices from these open tabs and save each as [vendor]-[date].pdf. First list which files you found, then ask me before downloading anything.
What to check: filename count and dates match the tabs. ChatGPT “will always ask for your permission” before managing downloads (per ExplainX’s coverage of the permission layer), so the ask-before-downloading instruction should be redundant — and still worth keeping.
The safety rules that make these prompts work
The extension’s Chrome permissions are broad — read and change data on all websites, browsing history, downloads, tab groups — so the prompts only pay off inside guardrails.
- Read-only first. Run the five Chat prompts for a day before granting Work mode access to anything that matters. That first day teaches you how the permission prompts behave on a site you don’t care about.
- Close sensitive tabs first. Banking, escrow portals, medical records, anything with card details — close them before Work mode runs. The extension requests access per new site, but the safest permission is a closed tab.
- “Stop before submitting” is not optional. Add it to every action prompt. OpenAI’s own sample includes it because independent reporting couldn’t confirm an approval fires before every submission.
- Know the permission tiers. “Allow for this site” is the sensible default; “Allow for all sites” is explicitly labelled Elevated Risk, and browser-history access is also Elevated Risk with no always-allow option (per ExplainX).
- Check your plan. OpenAI’s thread says all plans; PiunikaWeb reports Work mode needs at least Plus. Verify on your own account before promising the workflow to anyone.
- Enterprise users, check IT first. Managed accounts may carry administrator restrictions on Computer Use — if the option is missing, it’s probably policy, not a bug.
Related reading: our setup guide for ChatGPT’s browser agent walks through enabling Computer Use and the @Chrome tag step by step, and our Retriever AI review covers a free alternative that runs browser tasks without an OpenAI subscription.
Frequently Asked Questions
Is the ChatGPT Chrome extension free?
OpenAI’s launch thread says the update is “available for all ChatGPT plans,” and the extension itself is a free Chrome Web Store download (around 7 million users per the store listing, via Unite.AI). The caveat: PiunikaWeb reports Work mode — the part that acts — needs at least ChatGPT Plus. Check your own account before relying on it.
Can ChatGPT see my banking tabs?
It can read the tabs you grant it access to, and it requests permission for every new site. The honest answer is that nothing stops you from pointing it at a banking tab, which is why the advice from launch coverage is blunt: close sensitive tabs before you run Work mode.
What’s the difference between Chat and Work mode?
Chat mode reads: it looks across your open tabs and gathers information to answer, compare or summarise. Work mode acts: it fills out forms, sends messages, downloads files and organises tabs — with permission prompts along the way, though independent reporting couldn’t confirm every action asks first.
Does it work on browsers other than Chrome?
Per Unite.AI’s coverage, Edge, Brave, Opera and Vivaldi are supported via the ChatGPT desktop app — though Opera doesn’t get the side-chat panel. Chrome is where the full experience lives.
Should I let it submit forms and send messages by itself?
No. Keep the final click for yourself. OpenAI’s own sample prompts tell the agent to “stop before submitting,” and every Work prompt in this article carries that rule. Treat it like a fast assistant on day one: helpful, quick, and not yet trusted with “send.”
Sources and methodology
Launch details rest on OpenAI’s 15 October 2026 X launch thread as reproduced by PPC Land / 9to5Google, with setup, permission and plan caveats drawn from Unite.AI, ExplainX, and Gadget Review, all published 15–16 October 2026 and verified 21 October 2026. Seven of the prompts are adapted from OpenAI’s own launch-thread samples; three (research triage, tab cleanup, batch downloads) are original to this article. Nothing here was hands-on tested against a live install — extension behaviour varies by plan and region, and the Plus-requirement conflict is labelled as unconfirmed where it appears.
Govind Dheda covers AI tools, guides and prompts for OpenAIMaster — what is new, what is worth using, and how to put AI to work.