CAD-like Prompting / CADUI: A Judgment-Structured Framework for Human–AI Collaboration

Large language models are increasingly used not only for single-turn question answering but also for extended human–AI collaboration involving iterative drafting, analysis, revision, planning, and decision support. In such interactions, an important problem is not merely whether the model produces an incorrect output. Human judgments, model-generated interpretations, constraints, approval states, delegation boundaries, and provisional decisions may gradually become embedded in conversational history. As a result, it can become difficult to determine where a judgment originated, whether it was approved, what higher-level objective constrains it, what authority was delegated to the model, and where the workflow should return when a problem is discovered. This paper proposes CAD-like Prompting, a judgment-structured approach to human–AI interaction. Rather than treating prompting primarily as the construction of structured instructions, CAD-like Prompting represents collaboration as a Judgment Model composed of hierarchical objectives, interpretations, constraints, states, delegation boundaries, output conditions, and recovery information. The broader operational framework, referred to here as CADUI, separates a Human Judgment Layer (HJL) from a Machine Execution Layer (MEL). The HJL retains human judgment authority, while the MEL transforms approved or provisionally approved judgments into machine-executable operations such as retrieval, comparison, transformation, calculation, and generation. The framework further distinguishes judgment approval from artifact status and introduces explicit Draft, Review, and Final states together with Recovery Points, defined as saved judgment states rather than locations in conversational history. Failure handling is correspondingly framed not as an expanding taxonomy of model errors but as the identification of the judgment position at which the workflow became inconsistent and the minimum point to which it should return. The proposed framework is conceptual and operational rather than a claim of complete safety, correctness, or optimality. Its purpose is not to automate judgment itself, but to preserve where judgment resides, what state it is in, and how human judgment remains connected to machine execution throughout iterative human–AI work.