My personal Claude subscription costs $20 a month. In South Africa, at the current exchange rate, that is around R350. I max it out constantly on my Android phone.
The economic constraint is a burden. It is also the best experimental design I have accidentally stumbled into.
When every token costs something you notice, the hypothesis has to be precise before you spend it. The prompt has to be scoped. The question has to be small and focused. The context has to be exactly what is needed and nothing more. Not because a best practice document told me so. Because a vague prompt means I run out before the work is done.
The hypothesis is this: the better the definition upfront, the fewer tokens needed, with a better outcome. Every session is an experiment. Every constraint is data. The $20 account is not a limitation of what I can build. It is a forcing function for how I think before I build.
I suspect this would scale interestingly. A controlled experiment at one of the major AI labs, measuring output quality against token spend across different levels of specification precision, would tell us something useful about the economics of AI-assisted work. Every word counts. And it might count differently when the model or the context changes.
For now I run the experiment on a $20 budget. The results are in the research programme, the papers, the posts, and the governance framework. Not bad for R350 a month.