Reasoning discipline for hard, multi-step agent tasks — task decomposition, self-verification that kills fake "done", and next-step decisions. Use for architecture decisions, high-stakes changes, debugging dead-ends, or whenever a task feels too big for one pass.
