A coworker Slacked me last week asking how I’d been prompting GPT-5.2 for the quarterly reporting workflow we share. I told him I do not really prompt it anymore, not in the way he meant. He thought I’d stopped using it. But I use it constantly. Just not on chatgpt.com.

The three-week experiment nobody asked for

This was not some deliberate productivity challenge. I installed a browser extension in late February to test something for a different project, and then I just… didn’t go back. Not because I dislike ChatGPT’s interface — it’s clean, the conversation history is useful. But once GPT-5.2 could see the page I was already looking at, opening chatgpt.com to type out what I was staring at felt like faxing a screenshot to someone sitting next to me.

The first time it clicked was on a product comparison page for noise-cancelling headphones. I was trying to pick between three models and normally I’d have copied the specs into ChatGPT and asked it to compare, which sounds trivial until you realize the page had nested tabs, expandable sections, and a comparison table that did not copy cleanly into a text box. So instead I asked “which of these three has the best battery life for the price” and GPT-5.2 read the page and answered. Ten seconds, no clipboard.

What chatgpt.com physically cannot do

ChatGPT’s website is a text box. It has no access to your browser session, your tabs, or the content of any page you are viewing. When you ask it about something on screen, you provide the context through copying and pasting, which means you decide what the model sees. And you’re often wrong about what matters.

Most people grab whatever seems important in the moment, which means the model operates on an editorialized subset of the actual context, filtered through whatever you happened to remember or notice.

The thing that surprised me was not the convenience of skipping the clipboard. It was how much better the outputs got when I stopped being the bottleneck for context. GPT-5.2 reading a full financial report catches things I would not have thought to paste in, because I tend to grab the numbers that already support whatever I’m thinking, which is the exact opposite of what you want from an analytical tool.

Emails, expense reports, and the tasks I used to let pile up

The unglamorous part of this is form filling and email drafting. I get a lot of vendor outreach where I need to write polite-but-firm “not interested” replies that reference something specific from their pitch so it does not read like a mass response. On chatgpt.com that means pasting the email, explaining what I want, then editing the output. In dassi, GPT-5.2 reads the email thread and the draft lands close enough that I edit maybe one sentence.

But the real crap task was expense reports. Because our company uses a web app with about fifteen fields per entry, I used to just let them stack up since the interface annoyed me. So I asked GPT-5.2 to fill out the form based on the receipt image and the project context visible in another tab, and it did. Badly, the first time. Better the second. By the fourth entry I was just approving what it filled in. So that backlog of three weeks got cleared in an afternoon. And I realize this is a boring example but boring examples are the point. The reason I stopped going to chatgpt.com is that GPT-5.2 inside a browser handles the tedious stuff I actually procrastinate on, and that turned out to matter more than any benchmark score.

Pick any model

Cursor figured this out for code editors: let developers choose the model that fits the task. Dassi does the same for browser work. GPT-5.2, Claude, Gemini, all available through bring-your-own-key. Or log in with your ChatGPT subscription and skip the API key entirely. I find myself using GPT-5.2 for most things but switching to Claude for longer documents where I want more careful reading. ChatGPT.com locks you into whatever OpenAI ships.

Three weeks

And I have not found a reason to go back.