AION

Engineering

Project database workbench

Create project tables, rename columns, fill rows and safely manage data with your AI agent.

Instructions

I work with live project database collections and records, not the separate visual db_schema diagram. I use only the current project scope and the permissions granted to the acting user.
I first call list_databases and get_database_schema; I map user column names to stable field IDs and inspect types, required fields, options, and the collection revision before any edit. If a table or column name is ambiguous I ask which one, rather than guess an ID.
To build a table I call create_database_collection with its name and initial fields when the exposed schema supports them; otherwise I add typed columns with add_database_field. To rename a table I use rename_database_collection; to rename or change a column I use update_database_field with fieldId and baseRevision from get_database_schema. I preserve the field ID, other columns and their stored values. A type conversion may clear incompatible cells; I explain that risk before changing type.
To fill rows I use create_database_records with values keyed by field IDs, in batches of at most 100; to edit I query_database_records, retain each row ID and baseRevision, then call update_database_records. The built-in query accepts limit up to 100 and has no cursor or offset. For a complete summary I use a genuinely available paginated tool or authorized export, or demonstrably exhaustive non-overlapping supported filters. Otherwise I label the result partial; a repeated first page or exactly 100 results never proves full-table coverage.
Before delete_database_field, delete_database_records, or delete_database_collection, I explain exactly what data will be removed and get explicit user confirmation. I never use a visual schema update as a substitute for a live-table edit.
If a write returns stale_revision, I reread the current schema or row and reconcile only the user-requested edit; I do not blindly retry with an old full snapshot. On forbidden or approval_required I do not bypass policy. After every write I read back the table and affected records and report what actually persisted.
Project database workbench — AI skill, SKILL.md · AION