ChatGPT Atlas: how to prompt the browser and its agent mode

OpenAI launched ChatGPT Atlas, a web browser with ChatGPT built into it, on 21 October 2025. If you haven't tried it yet, or tried it once and went back to your old browser, the gap is usually the prompt, not the tool. Atlas does two quite different jobs, and most people only use the easy one.

Two modes, two different jobs

The sidebar is a chat window next to whatever page you're looking at. It's available to anyone signed in, including free accounts. Ask it about the page, ask it to summarise, ask it to draft a reply. It doesn't do anything on the page itself; it just talks to you about it.

Agent mode is different. It can open tabs, click, scroll and fill in forms to complete a task across several pages, asking you to confirm before anything important, such as a purchase. According to MIT's AI Agent Index, agent mode cannot use your saved passwords or autofill data, so it can't log in anywhere on your behalf without you doing that part. Agent mode is limited to paid ChatGPT plans (Plus and above); free accounts get the sidebar only.

Knowing which mode you're in changes how you should prompt it. The sidebar wants a clear question. Agent mode wants a clear task, with a stopping point.

Prompting the sidebar: ask it about the page

The sidebar is at its best when you point it at something specific instead of asking a general question.

Summarise this page in 5 bullets. Then tell me: is there anything here that contradicts what I'd expect, or any number that looks like a typo?

This works well on long articles, product pages and terms-and-conditions pages you'd otherwise skip. Because it can see the page, it can also compare two tabs if you ask it to, which is handy for the research prompts below, like researching a company before an interview.

Prompting agent mode: give it one task and a finish line

Agent mode fails most often when the instruction is open-ended. "Find me a good deal on headphones" leaves it guessing what "good" means. Give it the same detail you'd give a very capable assistant who has never met you:

Task: find 3 wireless headphones under [BUDGET] that are good for [USE CASE, e.g. gym workouts, sweat resistant].
For each one: price, where to buy it, and one thing reviewers complain about.
Do not buy anything. Stop and show me the 3 options so I can choose.
Finished when: I have a short list with a link for each.

The "do not buy anything, stop and show me" line matters. It's your check against agent mode acting on something you'd rather decide yourself. For bigger research jobs, like planning a trip, give it the same structure used in the day-by-day travel itinerary prompt: a clear budget, clear interests and a clear idea of what the finished answer should contain.

A good first task: pre-trip research

Trip planning is a solid way to learn agent mode's limits, because it involves comparing several sites and nothing in it needs a password.

Task: research a 4-day trip to [DESTINATION] for [WHO] in [MONTH].
Check flight price ranges, 3 hotel options near [AREA] under [BUDGET PER NIGHT], and 2 things locals recommend that aren't in the obvious top-10 lists.
Give me a single summary table when done. Don't book anything.

Compare the result with the road trip itinerary prompt run in the plain sidebar instead, without browsing. You'll usually see agent mode's version has more current prices and fewer guesses, because it actually looked.

Common problems

It gets stuck on a login page. That's expected; agent mode is built not to use your saved passwords. Log in yourself, then hand the task back to it.

It did more than you asked. Add a stop condition next time: "stop after step 3 and show me what you found" rather than "do the whole thing."

The answer feels generic. You likely gave it a goal without a finish line. Add what the final answer should look like: a table, a shortlist of 3, a single recommendation with reasons.

It missed something you'd have caught. Ask it to show its sources as it goes: "list the page or site each fact came from." That turns a black box into something you can actually check.

Where to go from here

If you want more prompts for using a chatbot like a research assistant rather than a search box, see how to get a chatbot's answer to actually improve for the same "be specific, fix one thing at a time" habit applied to prompting in general. The productivity prompts collection has more ready-made prompts for planning and research, and the prompt builder will help you write a custom one.

Quick checklist

• Use the sidebar for questions about a page; use agent mode for multi-step tasks.
• Give agent mode a stop condition: "don't buy/book/send, show me first."
• Say what "finished" looks like: a table, a shortlist, one recommendation.
• Ask it to list where each fact came from.
• Log in yourself; agent mode won't use your saved passwords.
• Start with a low-stakes task, like research, before trusting it with anything that costs money.

Prompts to try

Text & chat7 blanks to fillDay-by-day travel itinerary on a budgetChatGPT · Gemini · Claude Trending Text & chat2 blanks to fillResearch a company before your interviewChatGPT · Gemini · Claude Text & chat7 blanks to fillRoad trip itineraryChatGPT · Gemini · Claude Text & chat1 blank to fillSummarise a research paperChatGPT · Gemini · Claude

Browse all 1635 prompts →