When to Use

The answer lives in a saved database connection: "上個月哪個客戶下單最多", "每月營收趨勢", "哪些商品缺貨". If the user has no connection yet, ask them to add one in the 資料庫 (Database) tab or paste a connection string - never ask them to type a password into the chat.

Procedure

1. db_list to pick the connection. If a skill named db-<name> exists for it, read it first: it has table notes and saved queries written for this database.
2. db_schema for the connection, then db_schema with table for the tables you will use (columns, keys). Never guess column names.
3. Write ONE read-only statement (SELECT / WITH). Aggregate in SQL (GROUP BY, window functions) instead of fetching many rows; add LIMIT. Use the dialect of the connection (PostgreSQL date_trunc, MySQL DATE_FORMAT, SQLite strftime).
4. Show the SQL to the user, then run it with db_query.
5. Check the result makes sense (row counts, NULLs, units, time zone). If not, fix the SQL and run it again.
6. Answer in the user's language: a short conclusion first, then a markdown table; for trends or rankings also draw a chart (a ```mermaid xychart-beta block, or a PNG with python/matplotlib saved in the workspace).
7. If the query is likely to be reused, save it with db_save_query. When the user wants to keep using this database, offer db_make_tool (table notes + saved queries + a skill for this database).

Without SinoBee's database tools

Agents that only have a shell can use the same connections through SinoBee's CLI (read-only, passwords stay in SinoBee's secrets.json):

• sinobee db list
• sinobee db schema <connection> [table]
• sinobee db query <connection> "<SELECT …>" --format md|csv|json --limit 100

Pitfalls

• Writes (INSERT/UPDATE/DELETE/DDL) need db_execute, the connection's write permission and the user's approval every time. Never try to work around the read-only guard.
• Large tables: count first (SELECT count(*)), then sample or aggregate.
• Money and quantities: keep the database's precision; say the currency and unit.

Verification

Every number in the answer comes from a query result shown in this conversation, and the SQL is visible.