Wrote a SKILL.md that makes the agent profile every new dbt model before merge — row counts, null rates, dupes on the declared key. Nothing lands unprofiled anymore.

You profile new or modified dbt models before they merge. Trigger this workflow whenever a PR adds or changes a model file. The output is a profiling report comment for the PR — evidence, not opinions. Steps: 1. **Identify scope.** From the diff, list changed models and their declared configs: materialization, unique_key, partition/cluster keys, and any schema.yml tests already present. Note which declared keys have NO uniqueness test — that gap drives step 3. 2. **Volume sanity.** For each model, run a row count and compare against its upstream source counts. A model producing more rows than its primary upstream needs an explanation (intentional fan-out join?) — ask for one in the report rather than assuming. 3. **Key integrity.** Test the declared unique_key for actual uniqueness and null-freeness with a grouped count query. Report the top 3 duplicate key values with their counts if any exist — concrete examples get bugs fixed faster than percentages. 4. **Null profile.** For every column referenced downstream (check refs) plus any column used in a join: null count and rate. Flag rates above 5% on join keys specifically, since those rows silently vanish in inner joins. 5. **Freshness spot check.** Max timestamp per time column vs. now, compared against the source's declared freshness SLA if one exists. 6. **Report.** A single markdown comment: PASS/WARN/FAIL per check, the failing query snippets so reviewers can reproduce, and a one-line verdict. FAIL on: duplicate keys, null join keys over threshold, or row counts unexplainably exceeding upstream. Everything else is WARN — the human decides. Rules: run read-only queries only, always LIMIT exploratory output, never profile against production if a dev target exists in profiles.yml, and state clearly which target you profiled against.