Use when writing T-SQL for business intelligence, analytics, or reporting. Includes building summary reports with GROUPING SETS, ROLLUP, and CUBE, writing time-series queries with date bucketing, creating pivot/unpivot transformations, generating tally/numbers tables for gap-filling, building running totals and moving averages with window functions, writing year-over-year comparisons, designing materialized views for dashboards, or producing CSV/JSON exports from SQL Server.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
Use when writing T-SQL for business intelligence, analytics, or reporting. Includes building summary reports with GROUPING SETS, ROLLUP, and CUBE, writing time-series queries with date bucketing, creating pivot/unpivot transformations, generating tally/numbers tables for gap-filling, building running totals and moving averages with window functions, writing year-over-year comparisons, designing materialized views for dashboards, or producing CSV/JSON exports from SQL Server.
SQL BI & Reporting
When to Use
Writing summary reports with subtotals across multiple dimensions
Building time-series queries (bucketing, gap filling, moving averages)
Pivoting row data into columns or unpivoting columns into rows
Generating date ranges or number sequences (tally tables)
Writing queries optimized for Power BI DirectQuery or SSRS
Year-over-year, cohort, or funnel analysis
When NOT to use: application schema design (table design, naming conventions, access control), query performance tuning (execution plans, index tuning, wait stats), or ETL pipeline design.
Multi-Level Aggregation
Three operators produce subtotals from a single GROUP BY โ choose based on your needs:
Operator
Produces
Use when
ROLLUP(A, B, C)
(A,B,C), (A,B), (A), ()
Columns form a hierarchy (year > quarter > month)
CUBE(A, B)
All 2^n combinations
Need every cross-dimensional combination
GROUPING SETS(...)
Exactly what you list
Need specific subtotals, not a formula
-- Hierarchical subtotals: year > quarter > grand total
SELECT
YEAR(OrderDate) AS OrderYear,
DATEPART(quarter, OrderDate) AS OrderQtr,
SUM(TotalAmount) AS Revenue
FROM Sales.Orders
GROUP BY ROLLUP(
YEAR(OrderDate),
DATEPART(quarter, OrderDate)
)
ORDER BY OrderYear, OrderQtr;
Use GROUPING(col) to distinguish subtotal NULLs from real NULLs โ returns 1 for subtotal rows, 0 for data rows.
Full reference:Aggregation Patterns โ GROUPING_ID, conditional aggregation, HAVING vs WHERE, COUNT semantics with NULLs.
Window Functions
Quick Reference
Category
Functions
Frame needed?
Ranking
ROW_NUMBER, RANK, DENSE_RANK, NTILE
No
Offset
LAG, LEAD, FIRST_VALUE, LAST_VALUE
Yes for LAST_VALUE
Aggregate
SUM, AVG, COUNT, MIN, MAX
Yes for running/sliding
Running total
SELECT
OrderDate,
TotalAmount,
SUM(TotalAmount) OVER (
ORDER BY OrderDate
ROWS UNBOUNDED PRECEDING
) AS RunningTotal
FROM Sales.Orders;
Always specify ROWS explicitly. The default frame (when ORDER BY is present but no frame is specified) is RANGE UNBOUNDED PRECEDING โ this includes ties, produces different results, and is slower.
Moving average
-- 7-day moving average
AVG(DailyRevenue) OVER (
ORDER BY OrderDate
ROWS BETWEEN 6 PRECEDING AND CURRENT ROW
) AS MovingAvg7Day
Period-over-period comparison
-- Year-over-year monthly revenue (offset 12 for monthly data)
LAG(Revenue, 12) OVER (ORDER BY OrderYear, OrderMonth) AS PriorYearRevenue
Greatest-N-per-group
-- Most recent order per customer
WITH Ranked AS (
SELECT *,
ROW_NUMBER() OVER (
PARTITION BY CustomerID
ORDER BY OrderDate DESC
) AS RN
FROM Sales.Orders
)
SELECT * FROM Ranked WHERE RN = 1;
Named WINDOW clause (2022+)
When multiple window functions share the same partition and order, avoid repetition:
SELECT
OrderID,
ROW_NUMBER() OVER Win AS RowNum,
SUM(TotalAmount) OVER Win AS RunningTotal,
LAG(TotalAmount) OVER Win AS PrevAmount
FROM Sales.Orders
WINDOW Win AS (PARTITION BY CustomerID ORDER BY OrderDate);
-- GOOD: index seek
WHERE OrderDate >= '2024-01-01' AND OrderDate < '2025-01-01'
-- BAD: index scan (function on column)
WHERE YEAR(OrderDate) = 2024
Gap filling
LEFT JOIN from a continuous date spine (tally table or calendar table) to your sparse data. ISNULL replaces NULL with zero for missing dates. Use a range join to keep the date column SARGable โ never CAST(col AS DATE) in a JOIN:
SELECT
C.FullDate,
ISNULL(SUM(O.TotalAmount), 0) AS Revenue
FROM dbo.Calendar C
LEFT JOIN Sales.Orders O
ON O.OrderDate >= C.FullDate
AND O.OrderDate < DATEADD(day, 1, C.FullDate)
WHERE C.FullDate >= '2024-01-01'
AND C.FullDate < '2025-01-01'
GROUP BY C.FullDate
ORDER BY C.FullDate;
Full reference:Time Series โ calendar table design, fiscal calendars, cohort analysis, temporal table AS OF queries, moving averages.
Tally Tables
Generate number sequences with zero I/O using the stacking CTE pattern:
WITH
L0 AS (SELECT 1 AS c UNION ALL SELECT 1),
L1 AS (SELECT 1 AS c FROM L0 CROSS JOIN L0 AS B),
L2 AS (SELECT 1 AS c FROM L1 CROSS JOIN L1 AS B),
L3 AS (SELECT 1 AS c FROM L2 CROSS JOIN L2 AS B),
L4 AS (SELECT 1 AS c FROM L3 CROSS JOIN L3 AS B),
Nums AS (
SELECT ROW_NUMBER() OVER (ORDER BY (SELECT NULL)) AS N
FROM L4
)
SELECT N FROM Nums
WHERE N <= @Count; -- REQUIRED: limits output to needed rows
Every tally CTE must include WHERE N <= <limit> โ without it, the full 65,536 rows are generated. When used for date spines, the limit is DATEDIFF(day, @Start, @End). Never omit the limit and rely on an outer query to filter โ the optimizer may still materialize all rows.
On SQL Server 2022+, use GENERATE_SERIES(1, @Count) for simpler syntax. Do not use recursive CTEs for number generation โ they are row-by-row and 10-50x slower.
Full reference:Tally Tables โ inline function wrapper, date range generation, gap filling, string splitting, permanent vs inline tradeoffs.
Pivot & Unpivot
Static PIVOT (known columns):
SELECT ProductName, [Q1], [Q2], [Q3], [Q4]
FROM (...) AS Src
PIVOT (SUM(Revenue) FOR Qtr IN ([Q1],[Q2],[Q3],[Q4])) AS Pvt;
Detecting consecutive sequences (islands) and breaks (gaps) in ordered data.
Island detection โ the ROW_NUMBER difference technique:
-- GroupKey is constant within each consecutive run
DATEADD(day,
-ROW_NUMBER() OVER (PARTITION BY SensorID ORDER BY ReadingDate),
ReadingDate
) AS GroupKey
Group by GroupKey to find each island's start, end, and length.
Gap detection โ LEAD to find the next value, then check for breaks:
LEAD(OrderDay) OVER (ORDER BY OrderDay) AS NextDay
-- Gap exists when NextDay - OrderDay > 1
Full reference:Gaps and Islands โ session analysis, status change tracking, date-based vs sequence-based patterns.
Data Export
Format
T-SQL
Notes
JSON
FOR JSON PATH
Dot-notation aliases control nesting: AS "customer.name" produces {"customer":{"name":"..."}}
XML
FOR XML PATH
Only when downstream requires XML (SOAP, EDI)
CSV column
STRING_AGG(col, ',')
WITHIN GROUP for ordering (2017+)
Bulk file
BCP / BULK INSERT
TABLOCK for columnstore direct-path
FOR JSON PATH nesting
Use dot-notation aliases to control JSON structure โ no subqueries needed for flat nesting:
SELECT
O.OrderID AS "id",
C.CustomerName AS "customer.name",
C.Email AS "customer.email"
FROM Sales.Orders O
JOIN Sales.Customers C ON C.CustomerID = O.CustomerID
FOR JSON PATH;
-- {"id":1001,"customer":{"name":"Acme","email":"info@acme.com"}}
Full reference:Export Patterns โ FOR JSON nesting, JSON_OBJECT (2022+), Power BI DirectQuery optimization, BCP parameters.
Validation Checklist
Before finalizing any BI query, verify:
Subtotal rows โ GROUPING(col) = 1 filters identify rollup rows; confirm NULL isn't confused with real data
Window frames โ every windowed aggregate has an explicit ROWS BETWEEN clause
Date filters โ all WHERE and JOIN predicates on date columns use range predicates, never CAST(col AS DATE) or YEAR(col)
Gap fills โ LEFT JOIN from calendar/tally produces rows for every expected period; ISNULL handles missing data
Dynamic PIVOT โ generated SQL is printed and inspected before sp_executesql; all column names pass through QUOTENAME
Export shape โ FOR JSON PATH output matches the consumer's expected schema; test with a LIMIT before full run
Common Mistakes
Mistake
Fix
Default window frame with ORDER BY (RANGE, not ROWS)
Always specify ROWS BETWEEN ... explicitly
Using RANGE for moving averages (includes ties)
Use ROWS BETWEEN N PRECEDING AND CURRENT ROW
YEAR(col) = 2024 in WHERE (kills seeks)
Range predicate: col >= '2024-01-01' AND col < '2025-01-01'
CAST(col AS DATE) in JOIN conditions
Use range join: ON col >= date AND col < DATEADD(day, 1, date)
Tally CTE without WHERE N <= limit
Always add WHERE N <= @count โ without it, generates full 65K rows
COUNT(column) when you want total rows
COUNT(*) includes NULLs; COUNT(col) excludes them
AVG ignoring NULL semantics
AVG uses COUNT(col) as denominator โ use ISNULL(col, 0) if NULLs mean zero
UNPIVOT dropping NULL rows
Use CROSS APPLY VALUES to preserve NULLs
NOT IN with nullable subquery
Use NOT EXISTS โ NOT IN silently returns nothing when subquery contains NULL
Recursive CTE for number generation
Use the stacking CTE pattern โ set-based and 10-50x faster
FOR JSON AUTO in production
Use FOR JSON PATH โ AUTO changes shape when aliases change
HAVING for non-aggregate filters
Move to WHERE โ it filters before grouping, which is cheaper
FORMAT for date truncation
FORMAT is CLR-backed (10-50x slower) โ use DATETRUNC (2022+) or DATEADD/DATEDIFF
LAST_VALUE without explicit frame
Default frame ends at CURRENT ROW โ use ROWS BETWEEN CURRENT ROW AND UNBOUNDED FOLLOWING
LAG with fixed offset on sparse data
If periods are missing, LAG(val, 12) skips to the wrong year โ use a self-join
