Imagine teaching an AI to write code that not only functions correctly but also adheres to strict guidelines. This ability is crucial for businesses using domain-specific languages (DSLs) like JSON and YAML, where precise data structures and configurations are paramount. Researchers have explored this challenge in a new paper, "ConCodeEval: Evaluating Large Language Models for Code Constraints in Domain-Specific Languages." The paper introduces two key tests: generating valid data samples (Data as Code generation) and verifying existing code against a schema (DSL validation). The team tested various large language models (LLMs), including Llama and Granite, across five schema representations (JSON, YAML, XML, Python, and natural language). The results reveal that LLMs face significant hurdles in understanding code constraints, particularly those embedded within Python schemas, despite Python being a dominant language in their training data. Surprisingly, natural language proved to be the most easily understood for generating code samples. However, this didn't translate to the validation task. One intriguing finding was that models struggled when the schema and output code were in the same language. This research highlights a crucial area for improvement in LLMs: their ability to reliably work within the constraints of DSLs. This will be essential for broader AI adoption in tasks like system configuration, data exchange, and automated code generation. The next steps involve examining how LLMs reason about code constraints, expanding the range of constraints tested, and incorporating more complex scenarios, like understanding coding style preferences within natural language prompts and schemas. This research sets the stage for a new era of AI-driven code development where AI assistants can reliably generate and validate code within the specific constraints of diverse programming environments.