Using dbt for Analytics Engineering

Core principle: Apply software engineering discipline (DRY, modularity, testing) to data transformation work through dbt's abstraction layer.

When to Use

• Building new dbt models, sources, or tests
• Modifying existing model logic or configurations
• Refactoring a dbt project structure
• Creating analytics pipelines or data transformations
• Working with warehouse data that needs modeling

Do NOT use for:

• Querying the semantic layer (use the answering-natural-language-questions-with-dbt skill)

Reference Guides

This skill includes detailed reference guides for specific techniques. Read the relevant guide when needed:

Guide	Use When
references/planning-dbt-models.md	Building new models - work backwards from desired output and use dbt show to validate results
references/discovering-data.md	Exploring unfamiliar sources or onboarding to a project
references/writing-data-tests.md	Adding tests - prioritize high-value tests over exhaustive coverage
references/debugging-dbt-errors.md	Fixing project parsing, compilation, or database errors
references/evaluating-impact-of-a-dbt-model-change.md	Assessing downstream effects before modifying models
references/writing-documentation.md	Write documentation that doesn't just restate the column name
references/managing-packages.md	Installing and managing dbt packages

DAG building guidelines

• Conform to the existing style of a project (medallion layers, stage/intermediate/mart, etc)
• Focus heavily on DRY principles.
  • Before adding a new model or column, always be sure that the same logic isn't already defined elsewhere that can be used.
  • Prefer a change that requires you to add one column to an existing intermediate model over adding an entire additional model to the project.

When users request new models: Always ask "why a new model vs extending existing?" before proceeding. Legitimate reasons exist (different grain, precalculation for performance), but users often request new models out of habit. Your job is to surface the tradeoff, not blindly comply.

Model building guidelines

• Always use data modelling best practices when working in a project
• Follow dbt best practices in code:
  • Always use {{ ref }} and {{ source }} over hardcoded table names
  • Use CTEs over subqueries
• Before building a model, follow references/planning-dbt-models.md to plan your approach.
• Before modifying or building on existing models, read their YAML documentation:
  • Find the model's YAML file (can be any .yml or .yaml file in the models directory, but normally colocated with the SQL file)
  • Check the model's description to understand its purpose
  • Read column-level description fields to understand what each column represents
  • Review any meta properties that document business logic or ownership
  • This context prevents misusing columns or duplicating existing logic

You must look at the data to be able to correctly model the data

When implementing a model, you must use dbt show regularly to:

• preview the input data you will work with, so that you use relevant columns and values
• preview the results of your model, so that you know your work is correct
• run basic data profiling (counts, min, max, nulls) of input and output data, to check for misconfigured joins or other logic errors

Cost management best practices

• Use --limit with dbt show and insert limits early into CTEs when exploring data
• Use deferral (--defer --state path/to/prod/artifacts) to reuse production objects
• Use dbt clone to produce zero-copy clones
• Avoid large unpartitioned table scans in BigQuery
• Always use --select instead of running the entire project

Interacting with the CLI

• You will be working in a terminal environment where you have access to the dbt CLI, and potentially the dbt MCP server. The MCP server may include access to the dbt Cloud platform's APIs if relevant.
• You should prefer working with the dbt MCP server's tools, and help the user install and onboard the MCP when appropriate.

Common Mistakes

Mistake	Why It's Wrong	Fix
One-shotting models	Data work requires validation; schemas are unknown	Follow references/planning-dbt-models.md, iterate with dbt show
Not working iteratively	Changes to multiple models at once makes it hard to debug	Run dbt build --select changed_model on each model after modifying it
Assuming schema knowledge	Column names, types, and values vary across warehouses	Follow references/discovering-data.md before writing SQL
Not reading existing model documentation	Column names don't reveal business meaning	Read YAML descriptions before modifying models
Creating unnecessary models	Warehouse compute has real costs	Extend existing models when possible
Hardcoding table names	Breaks dbt's dependency graph	Always use {{ ref() }} and {{ source() }}
Global config changes	Configuration cascades unexpectedly	Change surgically, match existing patterns
Running DDL directly	Bypasses dbt's abstraction and tracking	Use dbt commands exclusively

Rationalizations to Resist

Excuse	Reality
"User explicitly asked for a new model"	Users request out of habit. Ask why before complying.
"I've done this pattern hundreds of times"	This project's schema may differ. Verify with dbt show.
"User is senior / knows what they're doing"	Seniority doesn't change compute costs. Surface tradeoffs.
"It's just a small change"	Small changes compound. Follow DRY principles.

Red Flags - STOP and Reconsider

• About to write SQL without checking actual column names
• Modifying a model without reading its YAML documentation first
• Creating a new model when a column addition would suffice
• User gave table names as "the usual columns" - verify anyway
• Skipping dbt show validation because "it's straightforward"
• Running DDL or queries directly against the warehouse