From Natural Language to Executable Filter Code: LLM-Assisted Behavioral Customization in Trigger–Action Platforms

Trigger–Action Platforms (TAPs) are a widely adopted class of End-User Development systems that enable non-programmer users to automate interactions among online services and IoT devices through simple IF-THEN rules. Although advanced mechanisms such as conditional execution and data transformation increase expressiveness, they typically require programming expertise and remain inaccessible to many users, leaving a persistent gap between users' automation needs and what template-based interfaces support. This paper presents an LLM-assisted approach for behavioral customization in TAPs, enabling the generation of executable filter code from natural language intents. The approach grounds generation in platform-provided catalogs of admissible variables and methods, guiding models toward syntactically valid and deployable solutions, and supports an interactive workflow in which users iteratively refine automation behavior through natural language. We evaluate the approach through a mixed-method study combining automated experiments and a controlled user study. The automated evaluation, conducted on 351 real-world rules, assesses platform compliance and behavioral correctness using surface-level metrics and a semantics-aware analysis grounded in the Trigger-Action execution model. The best-performing configuration achieves Syntax rates exceeding .97, API scores around .70, and CodeBERTScore approaching .79. The user study, involving 85 participants and 510 tasks, indicates that LLM assistance, in a controlled study setting, helps users converge toward correct automations, with success rates above 95%, good perceived usability (SUS M = 69.9), manageable cognitive workload (NASA-TLX M = 24.9), and moderately high trust in the generated solutions (TOAST M = 5.1).