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
Anti-Patterns
Based On
Data-quality practice — the six DQ dimensions, dbt tests / Great Expectations / source-freshness, severity-tiered enforcement.
Example Trigger Phrases
- "Define DQ checks."
- "Catch bad data before it hits dashboards."
- "Set up monitoring for a dataset."
1---2name: data-quality-checks3description: "Design the data quality checks for a table or pipeline across the standard dimensions. Use when asked to add data quality tests, define DQ checks, catch bad data before it hits dashboards, or set up monitoring for a dataset. Produces a checks plan across completeness, validity, uniqueness, freshness, consistency, and accuracy — each with the rule, severity, and where it runs (dbt test / Great Expectations / SQL assertion)."4disable-model-invocation: true # skill-optimizer5metadata:6 source: https://skillmd.com/skills/slipcast-dev/data-quality-checks7---89# Data Quality Checks Skill1011Bad data quietly poisons dashboards and models until someone notices the number is wrong. The fix is12checks that fail loudly *before* that — across the standard DQ dimensions. This skill designs them for a13specific table/pipeline: the exact rule per dimension, its severity (block vs. warn), and where it runs14(dbt test, Great Expectations, or a SQL assertion), so quality is enforced, not hoped for.1516## Required Inputs1718Ask for these only if they aren't already provided:1920- **The table/pipeline** and what it represents (grain, key columns).21- **The columns that matter** — keys, required fields, enums, ranges, dates.22- **Freshness expectation** — how current the data must be.23- **Tooling** — dbt tests, Great Expectations, Soda, or raw SQL assertions.2425## Output Format2627### Data Quality Checks: `[table]`2829Checks organised by dimension — each with the **rule**, **severity** (🔴 block the pipeline / 🟡 warn), and **where it runs**:3031| Dimension | Check | Rule | Severity | Implement as |32|---|---|---|---|---|33| **Completeness** | required fields non-null | `not_null` on [cols] | 🔴 | dbt test |34| **Uniqueness** | grain key unique | `unique` on [key] | 🔴 | dbt test |35| **Validity** | values in allowed set/range | `accepted_values` / range | 🟡 | GE / SQL |36| **Freshness** | data is current | max(loaded_at) within SLA | 🔴 | dbt source freshness |37| **Consistency** | cross-field / cross-table | e.g. totals reconcile, FK exists | 🟡 | SQL assertion |38| **Accuracy** | matches a source of truth | reconcile vs. system-of-record | 🟡 | SQL assertion |3940**Notes:**41- **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.42- **Where to check** — at ingestion (catch early) vs. in the model vs. post-build; recommend per check.43- **On failure** — what happens (halt, quarantine rows, alert + continue) and who's paged.4445## Quality Checks4647- [ ] Covers the core dimensions (completeness, uniqueness, validity, freshness, consistency)48- [ ] Each check has an explicit rule and a severity (block vs. warn)49- [ ] Severity is disciplined — only truly critical checks block the pipeline50- [ ] Freshness has a measurable SLA, not "should be recent"51- [ ] Each check names where it runs and what happens on failure5253## Anti-Patterns5455- [ ] Do not block the pipeline on every check — alert fatigue makes people ignore the real failures; reserve 🔴 for critical56- [ ] Do not only test the happy path — the grain key, nulls, and freshness are where the real breakage hides57- [ ] Do not write checks with no failure action — a test that fails into the void changes nothing58- [ ] Do not skip freshness — stale data that looks fine is the most dangerous kind59- [ ] Do not check only one table in isolation — cross-table consistency (FKs, reconciliations) catches integration bugs6061## Based On6263Data-quality practice — the six DQ dimensions, dbt tests / Great Expectations / source-freshness, severity-tiered enforcement.6465## Example Trigger Phrases6667- "Define DQ checks."68- "Catch bad data before it hits dashboards."69- "Set up monitoring for a dataset."