Data Quality Checks Skill

Bad data quietly poisons dashboards and models until someone notices the number is wrong. The fix is checks that fail loudly before that — across the standard DQ dimensions. This skill designs them for a specific table/pipeline: the exact rule per dimension, its severity (block vs. warn), and where it runs (dbt test, Great Expectations, or a SQL assertion), so quality is enforced, not hoped for.

Required Inputs

Ask for these only if they aren't already provided:

The table/pipeline and what it represents (grain, key columns).
The columns that matter — keys, required fields, enums, ranges, dates.
Freshness expectation — how current the data must be.
Tooling — dbt tests, Great Expectations, Soda, or raw SQL assertions.

Output Format

Data Quality Checks: `[table]`

Checks organised by dimension — each with the rule, severity (🔴 block the pipeline / 🟡 warn), and where it runs:

| Dimension | Check | Rule | Severity | Implement as | |---|---|---|---|---| | Completeness | required fields non-null | not_null on [cols] | 🔴 | dbt test | | Uniqueness | grain key unique | unique on [key] | 🔴 | dbt test | | Validity | values in allowed set/range | accepted_values / range | 🟡 | GE / SQL | | Freshness | data is current | max(loaded_at) within SLA | 🔴 | dbt source freshness | | Consistency | cross-field / cross-table | e.g. totals reconcile, FK exists | 🟡 | SQL assertion | | Accuracy | matches a source of truth | reconcile vs. system-of-record | 🟡 | SQL assertion |

Notes:

Severity discipline — only block on checks that should stop the pipeline (a duplicated grain key, stale critical data). Over-blocking trains people to ignore alerts.
Where to check — at ingestion (catch early) vs. in the model vs. post-build; recommend per check.
On failure — what happens (halt, quarantine rows, alert + continue) and who's paged.

Quality Checks

[ ] Covers the core dimensions (completeness, uniqueness, validity, freshness, consistency)
[ ] Each check has an explicit rule and a severity (block vs. warn)
[ ] Severity is disciplined — only truly critical checks block the pipeline
[ ] Freshness has a measurable SLA, not "should be recent"
[ ] Each check names where it runs and what happens on failure

Anti-Patterns

[ ] Do not block the pipeline on every check — alert fatigue makes people ignore the real failures; reserve 🔴 for critical
[ ] Do not only test the happy path — the grain key, nulls, and freshness are where the real breakage hides
[ ] Do not write checks with no failure action — a test that fails into the void changes nothing
[ ] Do not skip freshness — stale data that looks fine is the most dangerous kind
[ ] Do not check only one table in isolation — cross-table consistency (FKs, reconciliations) catches integration bugs

Based On

Data-quality practice — the six DQ dimensions, dbt tests / Great Expectations / source-freshness, severity-tiered enforcement.