It started with an inbound. Someone wanted a website — four pages, nothing fancy. At that point I’d opened Lovable maybe once or twice, just to see what the fuss was about. So I said yes, and started building.
I want to be honest about where I was standing, because it matters. I was a techie. I knew what Git was. But I had never written code in anger, never pushed to GitHub, never deployed anything. Push, pull, branches, builds — I knew the words and none of the actions.
The Lovable months
This was around September and October. Lovable was kind, in a way — it hid all the scary parts. It had a direct Supabase integration, it auto-deployed, and I never had to touch a terminal. For someone who didn’t know how to deploy, that was everything.
But getting the design right was brutal. I’d describe a page and watch it build something that was almost there and not quite. I’d ask it to keep both sides symmetrical. Sometimes it listened, sometimes it didn’t. I was spelling out every small detail, every button, every section, over and over. A single page could eat a week.
I burned through about $200 of subscription doing this. The UI was rough and I was paying in patience.
Cursor, and learning to actually ship
So I tried Cursor. It was better — sharper, and I could push it to do whatever I wanted. More importantly, this is where I finally learned the real machinery: how to connect my code to GitHub, how to push, how to deploy. The stuff Lovable had been hiding from me, I now understood.
That alone changed things. I wasn’t renting someone else’s deploy button anymore.
The week that became an hour
Then Claude Opus 4.5 dropped.
The process that used to take a week — define every section, place every button, ship it, check it, fix it — collapsed into about an hour. Maybe ninety minutes. I’m not exaggerating for effect; that was the actual difference. The same work, a fraction of the time, and the output was better.
There was a catch. On Cursor I was finishing $200 of credits in ten days. I bought another $200 and that was gone inside the month. My spend was crossing four, five hundred dollars — on a project I’d quoted at ₹40,000, which is barely $500. I was spending more on tools than the client was paying me. (That’s its own story, and I’ll write it properly another day.)
So I moved to Claude — the CLI, back when it had just come out — and the usage actually made sense. I’ve experimented since: Kimi K2.5, MiniMax, Codex, a few others. They’re interesting. But for real work, I’ve stayed on Claude.
Where I am now
Six months in, I’m comfortable. I’ve built simple websites, automations, AI workflows, integrations between different AI systems — I’ve even dabbled with voice agents. Not all of it deep, but enough to know I can pick up almost anything now.
That’s the part that surprised me most. Learning got easy. When I don’t understand something, I ask the AI to explain it, and I keep asking until it clicks. The barrier between “I don’t know this” and “okay, I get it” basically disappeared.
The speed is real. But the speed isn’t the point — it’s what the speed lets you do for the person on the other end of that inbound. That’s the whole idea behind this studio. The fast stuff, brewed well.