Fix a broken SQL query with Claude without guessing
Hand it the schema, the query, and the error. The fix is faster when the model can see the table shape.
A SQL error rarely makes sense in isolation. Paste the failing query, the exact error message, and the relevant CREATE TABLE statements. Claude reasons about joins and types far better when it can see the columns.
State what you expected versus what came back. 'This returns zero rows but I know the user exists' is a hundred times more useful than 'it doesn't work.' The model uses that gap to narrow the cause.
Ask for the corrected query and a one-line explanation of the bug, not just code. If it was a NULL in a NOT IN list or a missing JOIN condition, you want to recognize it next time.
Run the fix against a copy of the data, not production. Even a correct-looking query can have side effects, and a SELECT that looks safe might be the start of a write you didn't intend.
When the query finally works, ask Claude to rewrite it readably. The first fix is often ugly; a clean version is easier to maintain and review.