The smart zone vs the dumb zone

• Matt's next problem after Grill Me was turning a good planning conversation into reliable code. 40:46
• He uses Dex Horthy's "smart zone" and "dumb zone" idea: adding more context makes every token compete for attention. 41:14
• Matt estimates the smart zone of frontier models at about the first 150,000 tokens, regardless of a larger advertised context window. 41:34
• The practical problem is how to divide work that exceeds that useful context range across sessions. 42:03
• Ralph loops aimed to stay inside the smart zone by making the smallest useful change toward a goal, then clearing context. 42:31
• Matt keeps persistent state in the codebase and file system, not in one overloaded model context. 42:31
• His skill flow turns a grilling session into a destination spec, then breaks that spec into tickets, one ticket per session. 42:56
• The intended workflow is plan during the "day shift" and let agents execute during the "night shift." 43:42
• The goal is fewer context switches for the human and longer blocks where agents can work unattended. 44:05

flowchart LR
  A[Grill] --> B[Spec]
  B --> C[Tickets]
  C --> D[One session per ticket]
  D --> E[Review]
  E --> F[Next ticket]