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. Our agent can already read pages and fill in forms. Now it works across several sites at once. In this episode we ask the agent to gather data scattered across several websites into one tidy table, with its sources, and then we check it together.
Step 1: When data is spread across many sites
Have you ever opened ten tabs at once to compare something, then copied the numbers one by one into a spreadsheet? Work like that is tiring, and copying mistakes are easy to make. A browser agent can take over the boring part: opening sites one by one, reading them, then organizing the results into one table for you. There are four signs of a good fit. First, the data is spread across several sites, and no single source is complete. Second, the data isn't available through a connector. Courier sites or community sites usually don't have connectors. Third, each site looks different. Some use a table, some need you to fill in a price checker first, and some hide it in a paragraph. Fourth, the data needs to be compared side by side to make a decision. Our example today: Dewi is choosing a courier for regular shipments to Jakarta. If your task is just looking up one number on one site, open it yourself. An agent pays off most when there are many sites and the data needs organizing.
Step 2: Design the table first
The most common mistake is sending the agent off to search without deciding what the result should look like. Dewi starts with a clear task: compare shipping rates from Bandung to Jakarta for three couriers. Then the list of sites. All three are fictional for this example. The words no other sites stop the agent from wandering where it wasn't asked to go. The most important part is the format: the same six columns for every courier, including a source column. With uniform columns, results from three very different-looking sites can be compared. Finally, the limits: don't sign in or place orders. If the data isn't there, write not found. Don't guess. That last sentence saves Dewi from made-up numbers. After receiving the task, the agent doesn't set off right away. It asks two questions: regular service or cargo, and how many kilograms to use if a site asks for a weight. Dewi answers briefly: use the cheapest service, and use ten kilograms for rate checks. Questions like these are a good sign. It's better for the agent to ask at the start than to fill a rate checker with a guessed number. Dewi also prepares an empty spreadsheet with the same six columns. It will be filled from the agent's work. The source column is placed on the far right on purpose. Every number in the table must be traceable to the page it came from. Without a source, a number can't be checked.
Step 3: The agent browses three sites
The agent starts with the first site. The right panel shows it's on site one of three. This site is the easiest, because its rates are in a table. Bandung to Jakarta is nine thousand rupiah per kilogram, with an estimate of one to two days. The agent also reads the one-kilogram minimum, and the optional insurance cost. It records the source and the page's update date. That date matters, because rates can change at any time. The second site is trickier. Its rates aren't written in a table. To see a price, visitors have to fill in a shipping-cost checker: the origin city, the destination city, and the weight. The agent fills it in: Bandung, Jakarta, ten kilograms, because this site is cargo with a ten-kilogram minimum. Filling in a search box like this is safe, because it doesn't order anything. The result is sixty-five thousand rupiah for ten kilograms. The agent works out the rate per kilo itself, six thousand five hundred, so it can be compared with the first site. Then the third site. Here, rates only appear after signing in. Following the limits, the agent doesn't sign in. It records not found and notes the reason. The red line shows that Dewi needs to know about this decision. From those three sites, we learn three agent responses. Pages with tables or plain text are simply read and recorded. Search boxes or price checkers may be filled in, because they only look up information and don't commit you to anything. Login or ordering pages are skipped and reported, unless you yourself decide otherwise. Write these three responses into your task's limits, and the agent will work with the same pattern on any site.
Step 4: Results in a spreadsheet and checking them
Here are the results in the spreadsheet. Row one, QuickShip: nine thousand per kilogram, one to two days, with its source. Row two, Nusa Express: six thousand five hundred per kilogram, cheaper but slower, with a ten-kilogram minimum. Row three, Thrifty Cargo: not found, because the rates need sign-in. A row like this is honest. It's far better than numbers that look complete but are guesses. Dewi does a spot check. She opens the first row's source herself and matches the numbers. They match. One spot check builds confidence in the whole table. The table alone doesn't answer Dewi's question. She gives a real scenario: twenty-five kilograms a week, four times a month. Claude calculates the monthly cost: six hundred fifty thousand rupiah with Nusa Express, nine hundred thousand with QuickShip. The difference is two hundred fifty thousand, weighed against delivery speed. Claude also reminds her that Thrifty Cargo needs checking by Dewi herself after signing in. The final decision is still Dewi's, with complete and honest data.
Step 5: Ground rules for collecting data
Before trying it yourself, keep these ground rules. First, take the public data you actually need for your decision. Second, don't collect people's personal data, such as sellers' phone numbers or customers' addresses from a site. Even if it's visible on the page, personal data is still protected. Third, respect site rules. Some sites forbid automated data collection. Don't ask the agent to open hundreds of pages at once. Fourth, record the collection date and source of every row. Today's rates may differ next month. Fifth, spot-check on the original site before the data is used for a decision or shared with others. With these five rules, the agent becomes a diligent research assistant, not a data vacuum that troubles site owners. The same pattern works for many jobs. Nadia gathers a list of December small business events from three community sites for her clients. Mr. Hendra gathers scholarship program requirements from three institutions for the students at Cerdas Tutoring. Laras compares booth rental prices at three shopping centers. Notice the columns: an official contact, not the manager's personal number. The key is always the same: design the columns, name the sites, write the limits, then spot-check. The screens in this video are simulations, and Claude in Chrome may look different because it's updated often.
Done!
We've asked the agent to gather rates from three very different sites, organize the results into a spreadsheet with sources, and checked them with a spot check and a real scenario. Throughout this session the agent stopped several times to ask for a decision. In the next episode we cover how to make that call: when to allow an action and when to refuse. See you there!
Practice exercises
Try these on your own after watching to lock in what you learned.
Gather one type of data from three sites using the columns you designed
Check one result row by hand on its original site
Add a date-collected column and repeat the collection next month
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
52 video classes1,276 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.
