Visual LLM-Assisted Metadata Schema Construction from Spreadsheets Through Structured Editor Operations

Research data are often prepared in spreadsheets, but the meanings of their columns and values must still be translated into reusable metadata schemas. Even with a graphical editor, this translation requires repeated decisions about field names, types, properties, and hierarchy. Large language models (LLMs) can assist with these decisions, but workflows that return a complete JSON Schema make it difficult to inspect individual changes in a visual editor. We present an interactive visual LLM workflow for constructing metadata schemas from spreadsheets through structured editor operations. The workflow derives a schema draft from workbook information and represents both LLM suggestions and manual edits in the same operation model. Users can inspect, accept, or reject individual field-level operations in a graphical user interface (GUI) without reading or editing JSON Schema. The same mechanism supports updating an existing schema from a revised workbook while retaining user-approved changes. Experiments on six real-world datasets show that the operation-based workflow achieves type correctness and target-update accuracy above 97% across the evaluated models. Using the same LLM and datasets, it leaves fewer fields requiring correction than direct complete-schema generation or MetaConfigurator. A user study further shows that LLM assistance reduces the mean schema-construction time by 69.9%. The workflow gives researchers and data managers a practical way to create and maintain machine-readable metadata schemas for sharing spreadsheet data in accordance with the findable, accessible, interoperable, and reusable (FAIR) principles.