Modern LLMs can generate entire systems from natural language, but upfront specifications are merely hypotheses—real constraints emerge through implementation. Design is discovered by writing code, not just reviewing it. LLMs serve as brainstorming partners to explore design space and, once vocabulary is established, as natural language interfaces to domain models.

Domain Driven Design (DDD) advocates building a shared Ubiquitous Language in code. A DSL constrains syntax to express domain concepts precisely. DSLs like Plant UML, Mermaid, Graphviz, SQL, and Kubernetes YAML work exceptionally well with LLMs because their constrained grammar reduces variation. A few in-context examples reliably produce correct syntax.

For autonomous agents, DSLs provide deterministic validators (parsers, type checkers) enabling generate-check-repair loops without human intervention. Errors surface at domain level—"you cannot select an action before choosing a client"—not as buried stack traces. The trade-off: upfront cost to design and maintain the language, payoff concentrated in well-factored, constrained DSLs backed by validators.

A practical example: generating diagram-rich PowerPoint presentations for teaching distributed systems. A YAML specification references Plant UML sequence diagrams with step markers (e.g., nodes athens, byzantium, cyrene), producing step-by-step slides automatically.