Vander Labs

Skills / Data

SQL for analysts

The right grain, proven against a total

Your AI turns a business question into correct SQL the way an analytics engineer would: the grain and definitions pinned first, exploration queries, CTE-structured logic, safe joins and dates, window functions where they fit, and validation against a known number.

By Vander Labs. A Vander Labs Original, written by the Vander Labs team with AI assistance. Works with Claude, ChatGPT, Cursor and Claude Code.

What your AI learns

  • Restates every question as a result grain before writing SQL
  • Labels join cardinality and prevents fan-out double counting
  • Validates the result against a total the business already trusts

How it works

  1. Restate the grain: One row per what, with which measures, for which range and filters.
  2. Explore the data: Row counts, keys, null rates, date ranges and status values, run before the real query.
  3. Build with CTEs: One step per CTE with its grain in a comment; no deep nesting.
  4. Join on the grain: Cardinality stated for every join, many sides aggregated first, left-join filters kept in the on clause.
  5. Dates and time zones: Business time zone before truncation, half-open ranges, a calendar series for empty periods.
  6. Window functions: row_number, lag and lead, running sums and dense_rank instead of self-joins.
  7. Aggregate carefully: count distinct where identity matters, conditional sums, nullif on every division, ratios of sums.
  8. Validate and document: Reconcile to a known total, check row counts, spot-check a group, add a header comment.

Limits and safety

  • Flags assumed definitions as assumed and asks for confirmation of what counts as active, paid or signed up.
  • Not for schema design or production application queries.

Topics: sql, analytics, postgresql, bigquery, window-functions