Seth's NPC classifier is a textbook failure that almost shipped. The frame he gave Claude was: write a function that scores a state tag against a category name. Claude returned a function that scores a state tag against a category name. The output is excellent inside that frame. Outside that frame , in the world where the event log says haunting when the category is hunting, where the error mode is semantic substitution rather than spelling drift , the frame is the wrong question.

If decides what the build is, decides how to assemble it from Claude's actions. Claude Code is not one tool. You can ask it to plan, to write, to read, to refactor, to run tests, to run scripts, to search, to commit. Each of those is a different task type with a different failure profile. A planning prompt produces a list; if the list is wrong, you waste a build. A writing prompt produces code; if the code is plausible-but-wrong, you waste a debug.

The move Seth made when he typed hunting and haunting , before any formal verification, before the test runner confirmed anything , is Plausibility Auditing. It is the capacity to hear the wrong note before the proof is ready. Michael Polanyi, who was a physical chemist before he became a philosopher, built a career on a single sentence: we can know more than we can tell (Polanyi, 1966).

The reader knows the five capacities. Chapter 6 introduces the tool that makes exercising them systematic: Gru.

The reader knows the five capacities. Chapter 6 introduces the tool that makes exercising them systematic: Gru.