Dbt Skill

dbt Skill for Claude

Comprehensive dbt guidance: project structure, modeling, testing, CI/CD, production patterns. Targets Snowflake and BigQuery. Beginner-friendly with progressive scaling.

When to Use

Activate when: creating/modifying dbt models, choosing materializations, structuring layers, setting up tests, implementing CI/CD, configuring sources/freshness, writing Jinja macros, reviewing dbt projects, making analytics engineering decisions.

Skip when: basic SQL syntax, warehouse admin, raw pipeline config (Fivetran/Airbyte), BI tool config.

Core Principles

1. DRY via ref()/source() -- never hardcode table names
2. Single Source of Truth -- each concept defined once; staging = entry point, marts = consumer interface
3. Idempotent Transformations -- dbt run twice produces identical results
4. Test Everything -- every model has at minimum PK uniqueness and not-null tests
5. Progressive Complexity -- start with views/tables, add complexity when volume demands it

Project Structure

dbt_project/
├── dbt_project.yml
├── packages.yml
├── profiles.yml                 # local only, not committed
├── models/
│   ├── staging/                 # 1:1 with source tables
│   │   └── <source>/
│   │       ├── _<source>__models.yml
│   │       ├── _<source>__sources.yml
│   │       └── stg_<source>__<entity>.sql
│   ├── intermediate/            # business logic, joins, pivots
│   │   └── <domain>/
│   └── marts/                   # business-facing tables
│       └── <domain>/
│           ├── _<domain>__models.yml
│           ├── fct_<entity>.sql
│           └── dim_<entity>.sql
├── macros/
├── tests/
│   └── generic/
├── seeds/
├── snapshots/
└── analyses/

Layer Decision Matrix

Materialization Decision Matrix

ref() and source() Rules

1. source() only in staging -- staging is the sole gateway to raw data
2. ref() everywhere else
3. Never skip layers -- marts must not ref() staging directly
4. Never hardcode schema names

SQL Style

• Leading commas, lowercase keywords, CTEs over subqueries
• Explicit columns in marts (no select *), final CTE named final
• 4-space indentation, one column per line

Source Configuration

Define sources in _<source>__sources.yml with loaded_at_field and freshness thresholds. Configure warn_after and error_after per table.

Warehouse Quick Reference

Common Commands

Gotchas

• ref() only in staging-and-above, source() only in staging — using source() in marts bypasses the staging contract and breaks lineage
• Incremental models without unique_key silently duplicate rows on re-runs — always set unique_key for merge strategy
• ephemeral models can't be tested directly or selected with dbt test --select — test via downstream consumers
• dbt run doesn't run tests — always use dbt build to get run + test in DAG order
• --full-refresh on a large incremental model can blow warehouse credits — scope with --select first
• Jinja whitespace control: {%- -%} vs {% %} — trailing whitespace breaks compile output silently
• packages.yml version ranges (>=1.0,<2.0) can pull breaking changes — pin exact versions in production

-- WRONG: source() in marts (skips staging layer)
SELECT * FROM {{ source('stripe', 'payments') }}
-- RIGHT: ref staging model
SELECT * FROM {{ ref('stg_stripe__payments') }}

-- WRONG: incremental without unique_key (duplicates on re-run)
{{ config(materialized='incremental') }}
-- RIGHT: always specify unique_key
{{ config(materialized='incremental', unique_key='payment_id') }}

Scope Constraints

• Advisory guidance for dbt project design; runs dbt commands only when the user asks
• Warehouse specifics assume Snowflake or BigQuery — adapt for other adapters
• Does not cover dbt Cloud account setup or warehouse-side schema/grant administration
• Assumes source tables exist; raw ingestion (Fivetran, Airbyte, etc.) is out of scope

Reference Files

Load on demand when detailed guidance is needed:

License

Apache License 2.0. See LICENSE file for full terms.

Copyright 2026 Daniel Song