Prompt-State Machines: Applying Interface-Led Architecture to Structure LLM Reasoning

This technical note defines Prompt-State Machines as a prompt-level application of Interface-Led Architecture (ILA) for structuring Large Language Model reasoning. The paper applies the Provider–Observer–Operator abstraction discipline to prompt design. It treats prompts not merely as natural-language instructions, but as structured pseudo-state machines that separate premise provisioning, ambiguity observation, reasoning operation, clarification, and fail-closed termination. The paper introduces a minimal Prompt-State Machine structure using Markdown / ROM-like state boundaries, including Provider, Observer, Operator, Clarify, and FailClosed states. It also presents a Skill-Set Provider pattern, where a Provider declares not only task materials but also available skill boundaries, allowed operations, required inputs, forbidden assumptions, output contracts, and fallback behavior. This paper is positioned as ILA Supplement 2. It does not extend the general theory of ILA itself, nor does it import AIKernel-specific Semantic IR, Composite Distance, or ReplayLog as required prompt-level constructs. Instead, it demonstrates how ILA's role-based abstraction discipline can be applied to LLM prompt space as a lightweight governance protocol. The English manuscript is the canonical version. The Japanese manuscript is included as a companion translation.