SuperLede

I’ve slept less this week trying to burn through my quota before the reset. GPT-5.6 Sol is dangerously tempting.

The useful distinction may be ambiguity, not "new project vs bug fixing." Max for open-ended architecture work, high when the task and acceptance criteria are already clear. Cost per successful task would be the metric I'd care about most.