Tokens Were Going to Waste — So I Started Making a Screen Every Week

The starting point of a design experiment born from leftover tokens

I paid for a Claude subscription every month, but tokens kept going to waste. I’d check thinking I’d used them all — still there. Then next month, reset. It felt like a waste.

“I feel like I need to start a project just to burn through tokens.”

That’s the line that opened a session one day.

As it turned out, Claude had a separate design token pool. Credits that refreshed weekly, separate from coding tokens. Which meant every week, I automatically had the opportunity to build one screen. Forced creativity, basically — if I didn’t use them, they’d disappear, so I had no choice but to make something every week.

This didn’t last forever. Eventually the coding and design tokens merged, and that forced weekly creation routine naturally ended. But during that period, my hands got in the habit of building things — and another thought crept in.

“What if instead of doing this by hand every time, I automated it?”

There was something that connected to this naturally. The Atlas app-creator I was thinking through had a flow where users say “I want to build a café website,” a gallery of example templates appears, and they pick from there. Which meant I needed a lot of templates. Tokens going unused, a place to use them — the conditions lined up.

At first I thought about pulling in free templates. But free templates are free for a reason. Constraints, limited styles. But if I had Claude Code build them directly, my design sensibility wasn’t there to guide it. I sat with the problem and landed on one answer.

“What if Claude makes them?”

On claude.ai, you say “build me a page like this” and you get reasonably solid HTML/CSS as an artifact. Use that as templates.

But I took it one step further. There’s no point just burning tokens by randomly generating things. “One good skill is worth 100 templates.” The real target was a pipeline you could keep pulling from once built well. And the real experiment was: can Claude design and improve that pipeline itself?

“Can it freely generate topics, auto-generate multiple template versions, and verify the results itself?”

It didn’t start from ambitious planning. What do I do with leftover tokens — that was the starting point. But that small question rolled further than expected.