Sounds like you’ve been bumping into the same wall and asking the same questions.
Since I switched to this layered setup, the amount of “cleaning up after the programmer” dropped a lot. The programmer‑AI has become much more obedient, and most of the real fights now happen with the architect‑AI instead.
You’re absolutely right that it’s hard to front‑load enough architectural thinking. That’s also why this project raced its way to version 12 so quickly – every time the architecture felt wrong, I forced myself to rethink and re‑document it, not just patch the code.
On the token/pricing side, I honestly haven’t thought too hard about the exact math. One way or another, we’re all just helping to pay Nvidia’s bills, right?
Since I switched to this layered setup, the amount of “cleaning up after the programmer” dropped a lot. The programmer‑AI has become much more obedient, and most of the real fights now happen with the architect‑AI instead.
You’re absolutely right that it’s hard to front‑load enough architectural thinking. That’s also why this project raced its way to version 12 so quickly – every time the architecture felt wrong, I forced myself to rethink and re‑document it, not just patch the code.
On the token/pricing side, I honestly haven’t thought too hard about the exact math. One way or another, we’re all just helping to pay Nvidia’s bills, right?