At most one email a week. Unfollow anytime with one click.
Transcript
The full narration of this video. The current sentence is highlighted while the video plays; click any sentence to jump there.
Show transcript
Introduction
Hi, and welcome back to Claude from Zero to Expert. We're starting a new session, Claude in the Browser. In this session Claude doesn't just read data; it works right on web pages alongside you. In this episode we get to know Claude in Chrome: what it is, how it differs from connectors, how it works, and the safety rules for trying it the first time.
Step 1: From connectors to the browser
Claude in Chrome is Claude working inside the Chrome browser. It can read the open page, click buttons, type into fields, and move between pages, at your request. This is what we call an agent: an AI that doesn't just answer, but works step by step until the task is done. Availability may differ by plan and region, so check Anthropic's official page. How is this different from the connectors in the previous session? First, how it works. A connector uses the app's official connection. Claude in Chrome works through the page itself, exactly like we do with a mouse and keyboard. Second, what it's good for. Connectors are for apps like Gmail and Drive. Claude in Chrome is for sites with no connector: registration forms, supplier portals, or review sites. Third, speed. A connector fetches data directly. A browser agent works step by step, so it's slower. Fourth, supervision. Connector permissions are set up front. A browser agent needs watching while it works, because it acts on pages that could contain anything. So the two complement each other. If the app has a connector, use the connector. If not, Claude in Chrome is the way forward.
Step 2: How a browser agent works
The way an agent works can be pictured as a four-step cycle. First, look: the agent reads the content of the page that's open. Second, think: based on your task, the agent decides the next step. For example, the newest reviews are further down, so it needs to scroll. Third, act: clicking, typing, scrolling, or opening a new tab. Fourth, report: every step is logged in the side panel. Then the cycle repeats until the task is done, or until the agent needs your approval. Here's what to remember. First, an agent usually works more slowly than a person, because it reads the page before every click. That's normal. Second, an agent can misread a page. A complicated layout, a pop-up ad, or look-alike buttons can mislead it, just as they mislead people. Third, the agent works with whatever accounts are signed in to the browser. If you're signed in to a site, the agent acts on your behalf there. That's the main reason we're careful. Fourth, the agent can be stopped anytime. You see every step it takes, and you stay in control.
Step 3: First task: summarizing customer reviews
A task for an agent is written like a normal prompt, with a few additions. First, the task: summarize Sunset Coffee's reviews from the last two weeks. Second, name the site clearly, so the agent doesn't wander off to other sites. Third, the result format: praise, complaints, and suggestions, each with its number of reviews. Fourth, limits specific to agents: read only, don't sign in, and don't reply to reviews. These limits decide how far the agent may go. Let's look at the first task. Laras wants to know what customers have said about Sunset Coffee in the last two weeks. She opens the Claude panel in Chrome and writes the task. Look at the dark panel on the right. The agent opens Sunset Coffee's page on the review site. This site is fictional, just an example. Then it sorts the reviews by newest, because Laras only asked for the last two weeks. The top four reviews are read: the afternoon queue, the neat cold brew, slow WhatsApp replies, and the beautiful hampers. The agent scrolls down and reads eleven more reviews that fall within the two weeks. The last line matters: the agent doesn't reply to reviews and doesn't sign in. Its task is only to read, and it stays within that limit. The result appears in the Claude panel. Fifteen reviews are summarized into praise, complaints, and suggestions, with the number of reviews for each point. Laras wants to check one complaint. She asks exactly when Putri's review was written. Claude gives the date and includes a link to the original review. Laras can read it herself. The habit of checking sources still applies, just as when using web search. Work that usually takes half an hour of scrolling and note-taking is done in a few minutes, and Laras now has two real improvements for her coffee shop.
Step 4: A healthy limit: stop at the login page
The second example is from Mr. Hendra at Cerdas Tutoring. He wants to enter his students in a science olympiad, and asks the agent to find the registration schedule, fee, and requirements. The agent opens the organizer's information page. Again, this site is fictional. It reads the registration period: November first to twentieth, online. Then the registration fee: seventy-five thousand rupiah per participant. Then the requirements: students in grades four to six, with a letter from a school or institution. Cerdas Tutoring needs to confirm whether a letter from a tutoring center is accepted. Then the agent sees the sign-in button to register. It doesn't press it. The red line in the log shows the agent stopping to ask for Mr. Hendra's decision. Signing in and registering are Mr. Hendra's own job. That example shows a healthy limit. The agent stops and asks when it meets a login page or a password request. You always type passwords yourself, never the agent. Second, a payment page, a shopping cart, or a card details field. Third, buttons that commit you: send, register, agree, or delete. Fourth, a request to download a file or install something. Fifth, odd instructions written on the page, such as text telling an AI assistant to do something. As in the connector session, page content is reading material, not commands. If the agent stops at one of these points, that's not a failure. It's a sign the agent is working correctly.
Step 5: Getting started safely
Before trying it yourself, keep these four rules. First, check Anthropic's official page for availability, plans, and installation. The steps can change, so don't rely on old tutorials. Second, use a dedicated work Chrome profile. The agent works in the same browser as the accounts that are signed in, so a personal profile with bank and shopping accounts shouldn't be used. Third, start with reading tasks on public sites without signing in, like the two examples we just saw. Fourth, watch the screen while the agent works. If its steps go off track, for example opening a site you didn't ask for, just stop it. In this session we climb a trust ladder slowly. The first rung, reading, is what we did today. Next, filling in web forms, with the rule of stopping before the submit button. Then collecting data from several sites into one table. Finally, deciding when to allow an action and when to refuse. The screens in this video are simulations, and Claude in Chrome may look slightly different because it's updated often.
Done!
We've met Claude in Chrome, seen how it differs from connectors, learned the agent's work cycle, tried two reading tasks, and set a healthy limit of stopping at the login page. In the next episode we climb one rung: letting the agent fill in a web form from data we provide, then stop before the submit button. See you there!
Practice exercises
Try these on your own after watching to lock in what you learned.
Ask Claude in Chrome to summarize one review or news page about your business
Ask the agent to gather three key facts from one event organizer's site
Note every time the agent stops to ask permission, and why you approved or refused
Ask the class material
Type a question or keyword; we find the part of this episode that explains it and jump the video there.
My notes
Notes are saved with the current video time and are only visible to you.
Questions & answers
Stuck on a step? Ask here; the instructor and other learners can answer.
No questions yet.
Your instructor
Cerdas Inovasi Digital Team
Software developers, BIM & 3D engineers, and AI practitioners
53 video classes1,283 episodes
We build software, BIM models, games and AI automations for clients every day, and turn that hands-on work into step-by-step classes: every step is shown on screen and comes with a PDF guide you can keep.
