Researchers propose a domain-specific language to make AI prompt programming more reliable and maintainable

Large language models are becoming embedded in software systems at scale, but their probabilistic nature creates real challenges for reliability and maintainability. Researchers have now proposed treating prompt programming as a dedicated software engineering discipline, rather than scattering LLM-related logic throughout general-purpose code.

The Problem with Current Approaches#

Today's practice embeds prompt concerns — context management, interaction logic, output validation — directly in application code, creating implicit systems that are difficult to analyze, debug, or evolve. This scattered approach makes it hard to reason about what's happening when an LLM fails or produces unexpected results.

Introducing LLM Scripting Language (LSL)#

The team proposes LLM Scripting Language (LSL), a domain-specific language (DSL) designed to structure LLM interactions as explicit, analyzable programs. LSL introduces core abstractions including:

• Interaction blocks — discrete units of LLM communication
• Context scopes — controlled information environments
• Output constraints — validation and formatting rules
• Control flow — deterministic logic separate from probabilistic behavior

Benefits for Software Engineering#

By externalizing prompt design into programmable artifacts, LSL supports several engineering disciplines: it makes interaction logic visible and verifiable rather than buried in code, reduces implicit reasoning and assumptions, and enables systematic debugging and reuse. The framework acts as cognitive scaffolding, helping teams reason about LLM behavior more rigorously.

Limitations and Next Steps#

LSL does not guarantee semantic correctness or factual accuracy — these remain open challenges. However, it provides a principled foundation for more structured, maintainable prompt-based systems by improving failure localization and interaction transparency. The framework represents a step toward treating AI-powered software as a proper engineering domain.