Pivot Analysis Planner Skill

Pivot tables answer "what's driving X" in seconds — when the layout matches the question and the data is shaped right. Most pivot frustration is one of those two: a layout assembled by dragging until something looks meaningful, or data that isn't tidy (merged cells, subtotal rows baked in, one-column-per-month) feeding a pivot that double-counts. This skill works backwards from the question: decompose it, shape-check the data, choose the layout deliberately, and plan the drill-down — because the first pivot is the start of the analysis, not its output.

What This Skill Produces

• The question decomposition — the vague ask ("analyze sales") turned into pivotable questions ("which product line drove the Q2 change, and is it price or volume?")
• The shape check — the data problems that break pivots, found before the pivot finds them
• The layout plan — rows/columns/values/filters per question, each placement with its reason
• The reading guide — what the pivot will show, the drill-down path, and the so-what test

Required Inputs

Ask for these if not provided:

• The question, pushed to specific: "analyze this" becomes 2–3 answerable questions — the decomposition is half the skill, and it needs the user's intent
• The data's shape — columns and a sample; tidy (one row = one record, one column = one variable) or the fixes list gets written first
• The comparison that matters — vs last period? vs plan? across segments? Comparisons decide the column dimension and whether calculated fields are needed

Framework: The Planning Rules

1. Question first, layout second: each pivotable question maps to a layout — "what drove the change" → dimension in rows, period in columns, delta readable across; "where is it concentrated" → dimension in rows, values sorted descending, running % if available. Dragging-until-interesting produces coincidences, not answers.
2. Tidy data or fix it first: pivots need one-row-one-record: unmerge cells, delete pre-baked subtotal rows (they double-count), unpivot month-columns into a date column (data-cleaning-pass handles the mess). Five minutes of shaping saves an hour of pivot mystery.
3. Values need their aggregation chosen, not defaulted: SUM for additive things, COUNT for events, AVERAGE almost never without a weight-check (the average-of-averages trap). Every value field's aggregation is a decision with a reason.
4. Filters are the scope contract: what's excluded (test rows, internal accounts, incomplete current month) gets decided and stated in the output — an unlabeled filtered pivot is how two people present different totals from one dataset.
5. Plan the drill, then the so-what: the summary pivot points somewhere; the plan names the next cut ("if the East drop is real, re-pivot East by product") and ends at the so-what test — a pivot finding that suggests no action or decision was trivia, and the reading guide says which findings would matter.

Output Format

Pivot Plan: [dataset] — question: [the real one]

Decomposition

[The ask → 2–3 pivotable questions]

Shape Check

[Tidy? · the fixes needed first, if any]

Layouts

Reading Guide

[What each layout will show · the drill path · the so-what test: which findings change a decision]

Quality Checks

• Every layout traces to a decomposed question
• The shape check ran before any layout advice
• Every value field's aggregation has a reason; average-of-averages is guarded
• Filters/exclusions are stated where the output will be shown
• The drill-down path and so-what test exist — the pivot is a step, not the deliverable

Anti-Patterns

• Do not drag until interesting — layouts follow questions or they follow noise
• Do not pivot over baked-in subtotal rows — the double-count classic
• Do not average averages — weight or don't
• Do not present a filtered pivot without its scope note — that's how meetings get two truths
• Do not stop at the summary — the first pivot locates the question; the drill answers it