Getting correct SQL from AI tools: supply the schema, SQL dialect, and query plan

AI assistants can generate SQL quickly, but the results often run without error while returning incorrect numbers. The article attributes this to missing context, since models do not see the database schema, the engine in use, or the business definitions behind terms like revenue or active. A join across one-to-many tables can silently double-count totals, and NOT IN fails when a NULL is present in the subquery, returning no rows. Dialect differences between PostgreSQL, MySQL and SQLite also change results without causing errors, such as integer division and NULL sort order. The author recommends exporting the DDL with keys and nullability, stating the engine and version, and saving that context in the repository for reuse.