Linguistic control in AI text generation: An accessible prompt-based approach targeting L2 Spanish absolute beginners

Generative artificial intelligence (AI) offers strong potential for developing customized second language learning materials and tools. However, generating texts for absolute beginners which require strict lexical and grammatical control remains a challenge. Although controlled text generation (CTG) techniques exist, they often require technical expertise and infrastructure, limiting accessibility for educators. This study evaluates, in the context of Spanish, a prompt-based approach that leverages large language models (LLMs) without fine-tuning or specialized tools. Prompts enforce linguistic constraints defined in two attachments: a categorized Spanish vocabulary list, and a set of example sentences illustrating approved Spanish grammatical structures organized by communicative function. Three variables were manipulated: AI model (ChatGPT-4o vs. Claude 3.5 Sonnet), prompt type (standard vs. extended, with constraint-enhancing techniques), and attachment format (rich-heavyweight vs. lightweight JSON). A secondary variable, text type (city descriptions, personal introductions, and dialogues), was also examined. A total of 720 texts were generated, 30 per condition. Measures included proportions of non-compliant lexical and grammatical items, user-perceived latency, and errors in vocabulary, grammar, and coherence. Model choice was the primary driver of constraint adherence, with Claude 3.5 Sonnet outperforming ChatGPT-4o. Extended prompts improved adherence across models. Attachment format showed no systematic effect on adherence, but JSON significantly reduced latency and response-time variability. Text type also influenced adherence, and error rates remained low. Findings offer educators a scalable, low-barrier solution for generating tailored beginner-level Spanish materials and AI-powered tools using LLMs, along with insights into how different design choices affect performance. This approach, transferable to other languages, provides a practical alternative to resource-intensive CTG techniques, addressing a critical gap in AI-assisted language education.