The "Free" Path

Before I paid for anything, I tried to build this entire game for free. Not “free trial” free — actually free, indefinitely, across three different tools. It took about four months of casual development to admit that wasn’t working.

Cursor: five requests, then radio silence

Cursor’s free tier was my starting point, since it was going to be my IDE regardless. The pattern was almost comedic in its consistency: sit down, make maybe five real requests — chat back-and-forth that turned into actual code changes — and burn through the entire month’s allotment in well under an hour.

Then nothing. I’d come back a month later, get another hour of real work done, and go quiet again. I was effectively building this game in one-hour sprints, once a month. At that rate the math doesn’t work — you can’t hold a project’s context in your head across a month-long gap and still move fast.

Claude free: copy-paste development

Next was the Claude free tier, using Opus through the chat interface. This meant no editor integration — I’d ask for a file or a snippet, get back code, and manually copy-paste it into my project. It’s a workflow that technically works but adds friction to every single change.

The token refresh was faster here — every six hours or so I’d get another batch — but “faster than once a month” still isn’t fast. Between the copy-paste tax and the wait between batches, this wasn’t a sustainable way to build anything of real size.

Gemini: great for lookups, dangerous for design

Gemini ended up filling a different role entirely — it was genuinely good for quick lookups, in a way that reminded me of Stack Overflow in its prime. (Those were the days.) Fast answers to narrow, well-defined questions.

Where it fell apart was anything requiring actual design judgment. It struggled to help with the harder architectural and UX questions, and at one point gave me a suggestion that would have rm’d major files in my project outright. I didn’t run it. That one stuck with me — it’s a good reminder that “AI wrote it” is not the same as “AI got it right,” and that the tool most willing to hand you a confident answer isn’t always the one that should.

Breaking down and paying for it

After about four months of this — stitched-together hour-long sessions, copy-pasting code by hand, and one very nearly catastrophic rm — I got more motivated to actually push this game forward, and the free-tier math stopped making sense. I upgraded to Claude Pro.

That single change did more for development speed than anything else in the previous four months combined — which, in hindsight, says less about Claude specifically and more about what a real, unthrottled feedback loop is worth when you’re building something this size solo.

Next up: some hard truths about what building a production-level game with AI actually takes, once the honeymoon period with the tools wears off.