answering-natural-language-questions-with-dbt_skill

Overview

This skill answers business questions by querying dbt-managed data using a prioritized, safe workflow: semantic layer first, then modified compiled SQL, then model discovery, and finally manifest/catalog analysis. It is intended for responding to user data requests (e.g., total sales last quarter) and not for developing, testing, or running dbt models. The skill always exhausts options before returning that a question cannot be answered.

How this skill works

First check for a semantic layer and attempt to resolve the metric via list_metrics, get_dimensions, and query_metrics. If the semantic layer needs minor changes, retrieve compiled SQL (get_metrics_compiled_sql), adjust the raw SQL, and execute it. If no semantic layer is available, discover mart models with get_mart_models and get_model_details and build SQL referencing models. As a last resort, inspect target/manifest.json and target/catalog.json (filtered) to infer schema and craft warehouse SQL.

When to use it

• Answering business questions that require querying the warehouse (sales, active users, revenue by dimension).
• When a semantic layer is available and you want a metric-first approach.
• When a metric exists but needs a small SQL modification (extra dimension, custom filter, different aggregation).
• When no semantic layer exists but model discovery tools are available.
• When in a dbt project without a server and you can read manifest/catalog artifacts.

Best practices

• Always prefer the semantic layer for correctness and reuse before writing raw SQL.
• If you modify compiled SQL, work with the resolved SQL (no Jinja) and suggest adding the change to the semantic model.
• Prefer mart models over staging when discovering models; marts contain business logic.
• Filter manifest/catalog files before reading to avoid large, slow operations.
• Do not recommend database schema or ETL changes—keep suggestions at the semantic layer level.

Example use cases

• User asks: 'What were total sales last month?' — query the semantic metric or run compiled SQL if needed.
• User requests revenue by region with a custom segment — get compiled SQL and add CASE/WHERE/grouping.
• No semantic layer: discover mart models, inspect columns, write SQL using ref('mart_model'), and execute.
• In a local dbt project without a server: jq-filter manifest.json/catalog.json to find column info and produce SQL.
• Metric exists but lacks a dimension: join the compiled metric to a dimension table and return grouped results.