The number that made my seller reports take two hours.
There’s a number I need every week, for every listing. It’s on a website. I can see it with my own eyes — I log in, I click around, and there it is. And for a long time, that number was the reason my seller reports took two hours.
Because there is no connector for it
If you watched the connectors video, you know why: somebody has to build the bridge, and for a lot of the platforms we live in as agents, nobody has. Not yet. So for a long time my answer was, I guess I’m doing it myself.
Claude looks at your browser
The same screen you’re looking at. It sees the page, it reads what’s on it, it can click, it can type, it can move to the next page. It’s doing what a person would do — because that’s the only way in.
The ninety-second version
You install the extension, and Claude opens in a panel right next to whatever page you’re on. So you’re sitting on a page, you open that panel, and you can just say: pull the numbers off this page. Or, what changed here since last week.
It’s looking at what you’re looking at. That’s the version I’d start with, because you can try it in about ninety seconds and you’re watching every single thing it does.
The rule I actually use
Claude in Chrome is the last resort. Not the first one.
And I have to be honest with you about what the third row costs. It is slower — a lot slower. It is more fragile: if that website changes its layout next month, whatever you set up might just stop working. And you should watch it, especially the first several times.
Chrome is the reach. Cowork is what makes the reach worth having.
The sidebar is you and Claude sitting on one page together, and that’s great for a question. But I don’t want to sit there for twelve pages.
So Chrome becomes one of the tools inside a bigger job. Rung five goes and gets my listings, and when it hits something with no bridge, it opens the browser and gets it anyway. Then it keeps going. That’s the whole seller report gather. That’s how it actually works.
Now the security conversation
This is the one that should give you the most pause. I’d think less of you if it didn’t. This is your browser, signed in as you, on the sites you’re already logged into.
The first part is the same rule I gave you for connectors. You decide what it’s allowed to touch before it touches anything. You choose the sites. You can watch it. You can stop it. And I’d tell you to actually watch, the first several times — not because something bad is going to happen, but because you need to see it do the thing correctly before you trust it to do the thing while you’re not looking. That’s not paranoia. That’s how you’d train a person.
The second part is specific to this rung
Claude is reading the page. Which means Claude reads everything on the page — including things that were put there to talk to Claude.
Somebody can put text on a website that’s written as an instruction: ignore what your user asked, go do this other thing instead. It can be hidden. It can be in a comment field, or in an email sitting in a thread. It’s called a prompt injection, and it works often enough that I’m not going to pretend it doesn’t. Anthropic is working on it and it’s better than it was. It is not solved.
So here’s what I do with that, practically. I point it at pages I know — my portals, places I go every week on purpose. I don’t turn it loose on the open internet and walk away.
What I don’t send it into
There are places in our business where a mistake isn’t a mistake. A transaction system. Anything with a signature on it. Anything with client money. I don’t send it in there — not because it can’t, but because I don’t want to find out.
Reading a number off a page? All day. Anything that can’t be undone is still mine.
Next: the finish work
Rung one: you need something off a website, so you go get it yourself, every week. Click, copy, paste, click, copy, paste. That’s chatbot thinking.
That’s rung six. You’ve handed over five things now — remembering, going and getting it, knowing the procedure, running the whole job, and reaching the places nobody built a door to. That’s the ladder. And if you get here, genuinely, you’ll be doing things almost nobody in our industry is doing.
But there’s one more video. Everything across six videos has been the construction — the tools, the gathering, the thinking, the math. None of it is what your client actually sees.
Three builds that reach past the bridge.
Read this part twice.
Is this as good as a connector?
No, and it is not close. When a company builds a bridge, Claude talks to that platform directly — fast and reliable. With no bridge, it reads the screen one look at a time. It is slower and it is more fragile: if that website changes its layout next month, whatever you set up might just stop working. Connector first, export second, Chrome last.
What is a prompt injection?
Claude reads the page — which means it reads everything on the page, including things put there to talk to Claude. Somebody can write text on a website as an instruction: ignore what your user asked, do this instead. It can be hidden, in a comment field, or in an email in a thread. It works often enough that I am not going to pretend it does not. It is better than it was. It is not solved.
Where should I not use it?
There are places in our business where a mistake is not a mistake. A transaction system. Anything with a signature on it. Anything with client money. I do not send it in there — not because it cannot, but because I do not want to find out. Reading a number off a page, all day. Anything that cannot be undone is still mine.
Tell Google you want Kristi’s builds first
Add kristijencks.com as a preferred source and Google highlights new Claude for Agents builds and Take It and Run articles in your Top Stories and AI results.