Prompt library
Audit data quality before anyone quotes it
Layered checks catch different failure classes: structural breaks are loud, semantic contradictions are quiet, and reconciliation against independent systems is the only check that catches a pipeline lying consistently. Thresholds and block-decisions convert quality from a feeling into an operating rule.
Last reviewed July 17, 2026
The prompt
Build the data-quality audit for the dataset below, sized to how the data will be used.
Dataset: {{dataset}}
Stakes (what decisions ride on it): {{stakes}}
Produce checks in four layers, each as a concrete test with a pass threshold:
1. Structural: row counts vs expectation, duplicate keys, schema drift since last period, referential orphans.
2. Statistical: null rates per column vs their historical band, cardinality jumps, distribution shifts on the 5 most decision-relevant columns (state which and why).
3. Semantic: cross-field contradictions (end before start, refunds exceeding purchase, status/timestamp mismatches), unit mix-ups, timezone consistency.
4. Reconciliation: the 2-3 numbers that must match an independent system (billing to bank, events to vendor dashboard), with acceptable tolerance stated.
Then: the run cadence for each layer given the stakes, what a failure at each layer blocks (publish? decide? nothing but a ticket?), and the single-page pass/fail format the audit reports in.
Every check needs a threshold; "looks reasonable" is not a check.Run in idaptOpens a new chat with the prompt prefilled. Nothing sends until you press send.
Fill in the variables
| Variable | What it is | Example |
|---|---|---|
| {{dataset}} | The dataset | the revenue mart feeding monthly reporting |
| {{stakes}} | What the data drives | board reporting and sales commissions |