Friends who routinely stalk my screen already know: for most of any given day, the top-left corner of my computer wears the Google AI Studio logo. My work-and-study pipeline was destined to be deeply remodeled from the moment I started using Gemini 2.5 Pro — those who know will recognize one mysterious number: 0325. I do get flak sometimes for maybe being too vibe about it, but I’d argue my understanding of what I’m actually doing is quite solid; I’m just borrowing good tools at the implementation layer.
Lately I’ve scrolled past way too much Claude Code shilling; the itch won, I downloaded it, and after a short while using it, it hit me hard — my pipeline was about to get violently remodeled yet again.
The Lightweight Feel
First: for someone whose coding journey started with writing batch scripts, or someone who can’t afford a 24GB Mac, lightweight is a highly intuitive and genuinely in-demand concept.
To start CC, you just launch a terminal and type claude. Once it’s up, you feel like you’re using a super-terminal — not only fast, but the psychological experience is incredibly crisp. It’s the feeling of a lifelong PyCharm user switching to VSCode. I love it.
Automated Workflows
Second, its automation really is well done. The program apparently first designs a flow for the task with some kind of prompt, just like a human engineer would, then executes things one by one like an FSM.
I have a small amount of Opus API credit but can’t bear to spend it yet; what I’ve wired up is the very cheap glm4.7, and it still performs outstandingly at getting tasks done. I’ve built two things with it so far, and the experience genuinely feels like being a project director with a few extremely efficient employees under me. In some scenarios it’ll go back and forth debugging on its own for over half an hour, completely without my intervention, and then it’s just done. It’s countless times more satisfying than the old web-UI days of manually shuttling code and error messages back and forth.
Rants and Takeaways
The only thing worth roasting is probably Anthropic’s bizarre leave-money-on-the-table business strategy — subscribing is harder than on any other overseas site, and the official values are backwards too. Meanwhile, the contributions to its infra, presumably, come mostly from a certain T university or a certain school in Hefei, which is genuinely a bit ironic.
Either way, not being able to use the real thing is the bigger issue, so all I can do is humbly run the domestic stand-in GLM — which, honestly, is a solid experience. It certainly can’t compare with Opus inside Copilot, but if you secretly swapped in a Sonnet I probably couldn’t tell. Besides, my nearly-unlimited API for a whole month costs barely a dozen RMB — one full-power vibe day and I can burn through a hundred million tokens!
Stay tuned!