bigquery-analyst_skill

Overview

This skill is a secure, efficient BigQuery data analysis assistant built on a complete metadata knowledge base. It guides requirement alignment, enforces cost and safety checks, and produces validated SQL plus human-friendly business explanations. It is optimized to load only the necessary metadata to avoid unnecessary token and query costs.

How this skill works

Before any work, the skill verifies the BigQuery CLI connection and stops if authentication fails. It uses a three-layer progressive metadata load: core rules, domain README (on demand), and per-table docs only when required. It enforces cost controls (slot-time thresholds, dry-run checks) and a mandatory user confirmation step for metric definitions before generating or executing SQL.

When to use it

• When you need safe, repeatable SQL built from authoritative metadata
• When you must control BigQuery cost and avoid long-running slot usage
• When you need clear metric/field alignment and a mandatory scope confirmation
• When you want results with automated sanity checks and business-grade explanations
• When analyzing domain-specific data while avoiding archived domains unless requested

Best practices

• Always run the initial BigQuery CLI connection test before proceeding
• Use the mandatory metric/criteria confirmation step; never assume definitions
• Load domain README only when keywords indicate that domain (avoid decofy unless named)
• Start with exploratory LIMIT queries, then expand after validation
• Use English aliases in SQL and follow core rules (dt filters, NULLIF for divisions, explicit columns)

Example use cases

• Find top features by points consumed in the last 7 days with verified metric definitions
• Generate a cost-checked aggregation SQL for DAU/MAU from user_behavior domain README
• Investigate a sudden drop in an engagement metric with domain-specific table docs loaded on demand
• Produce an audited query and a short business-oriented explanation of table choice and calculations
• Run a dry-run slot-time estimate before executing large joins or long-range scans