dbt Wizard vs Claude Code: same tasks, deeper checks, $3.17

The test put both agents on the same underlying model, claude-sonnet-5, so any difference reflects agent design rather than the model. dbt Wizard's edge comes from how it engages the project: before editing it checks upstream and downstream impact, it confirms dbt_utils was not installed before writing tests, and on the churn logic it did not hard-code the 90-day threshold, instead pulling it into a variable.

The contrast on the ambiguous churn task was the clearest. Wizard compared 90-day and 30-day thresholds against real data and reported that 90 days produced no churned customers with order history. Claude Code hard-coded 90 in the SQL. Claude Code shipped faster: roughly 23 to 58 seconds per task versus Wizard's 4 minutes 36 seconds on task 1 and 4 minutes 44 seconds on task 5.

The author's team already runs Claude Code with dbt rules in place, so their preference is to enrich Skills rather than pay for a second agent. Still, for teams without that setup, Wizard's default habits — checking git diff before fixes, verifying with dbt build, and testing whether a chosen threshold produces meaningful segments — stand out as practices Claude Code needed to be told to follow.