Judgment is the human reasoning used to decide what to build, assess evidence and risk, and determine whether a software change deserves to ship. AI can accelerate familiar implementation work, but the context-sensitive remainder still requires someone who understands the consequences. The “70% Problem” is shorthand for this divide, not a measured industry statistic. section 1.6 section 10.8.3
Fast generation can create an illusion of being nearly done. A component may handle the happy path yet fail on an unexpected state, or pass its own tests while violating the surrounding architecture. Faster agents can also propagate a wrong decision into other services before anyone notices. section 1.6 section 10.8.1
Apply judgment at several decision points:
- Question the goal: decide whether a service should exist, whether a requirement captures the user’s need, and whether an old architectural decision still fits. section 10.8.1
- Evaluate the change: build a mental model of the code, probe failure paths, and ask whether tests cover important behavior rather than merely easy cases. section 8.8.5 section 8.8.6
- Calibrate AI assistance: choose when direct retrieval is sufficient, when reasoning is worth its cost, and when to intervene and check. section 3.6.4
- Improve the feedback system: maintain shared context, truthful dependency information, honest tests, fast feedback, and review that scales. section 10.8.4
When less reasoning is enough
A model answered a question about a game’s initial state by recognizing a widely documented quirk. A direct retrieval succeeded without a reasoning chain or exploration loop. That does not make retrieval universally trustworthy; it illustrates why the choice of technique itself requires judgment. section 3.6.4
Checklists set a floor, not a ceiling. Passing tests and coverage numbers can give false confidence, especially when the specification is ambiguous or the tests are weak. Be explicit about what has been verified and what remains an assumption. section 8.8.5 section 8.1.4


