Data Governance Agent Skill

Data governance

Governance has a reputation for bureaucracy because it is usually implemented as approval queues.

Done properly it is the opposite: it makes data usable without asking anyone.

Start with definitions, not policy

The highest-value governance artifact is a metric dictionary. For each business metric:

• The plain-language definition — what it counts, and what it deliberately excludes.
• The computation, unambiguously: source table, filters, time grain, timezone.
• The owner — a person who decides when it is disputed.
• Known caveats — when it is misleading, and what changed historically.

Most metric disputes dissolve once both parties read the same definition and discover they were

measuring different things. Almost none require a policy.

Watch the ones that look obvious. "Active customer," "revenue," and "signup" each have half a dozen

defensible definitions, and the ambiguity surfaces at the worst moment.

Ownership

Every dataset has a named owner accountable for its quality and access — a person, not a team.

Unowned datasets decay, and nobody notices until a decision is made on stale data.

The owner should sit with the business meaning, not with the pipeline. The team that generates the

data understands what it means; the platform team understands how it moves.

Quality, measured rather than asserted

Test data like code, continuously, and alert on failures:

• Freshness — did it arrive when expected?
• Volume — is the row count within its normal range? A silent drop to zero is the classic

failure.

• Uniqueness and nullity on key fields.
• Referential integrity across joins.
• Distribution — has the shape shifted in a way nothing explains?

The point is finding breakage before a decision is made on it. A pipeline that fails loudly is

better than one that silently produces yesterday's numbers.

Access

Default to open for internal, non-personal data. Restrictive-by-default drives the shadow spreadsheet

layer, which is genuinely less safe than a governed warehouse.

Personal, financial, and regulated data are the exception: least privilege, purpose stated, reviewed

periodically, with Legal & Risk involved on anything with a lawful-basis question.

Lineage

Know where a number came from and what feeds it. Without lineage, you cannot answer the two

questions that matter during an incident: what broke upstream, and what downstream is now wrong.

Never

• Let two systems each claim to be the source of truth for the same fact.
• Fix a data-quality issue in a dashboard. Fix it upstream or it recurs in every other consumer.
• Retire a dataset because it looks unused — you cannot see every consumer. Deprecate, announce,

then remove.