---
name: "consulting_charts"
description: "Rebuilds data charts the consulting way: picks the chart by its message, strips non-data ink, annotates the takeaway, and frames slides around action titles. Use when making board-ready or executive slides, fixing charts that bury their point, or choosing between chart types for a presentation."
metadata: { "includeInPrompt": true }
---

# Consulting Charts

## Purpose
Turn a data dump into a chart an executive reads in ten seconds. Every chart carries exactly one message, chosen before any charting tool opens. The workflow: message first, chart second, ink discipline third, annotation fourth, slide narrative fifth.

## Workflow

### 1. State the message before touching a chart
Write the slide's action title first: one complete sentence, 15 words or fewer, active voice, containing the key number. Examples:
- "Enterprise grew 22% while mid-market shrank 4%, shifting the mix"
- "Northeast region drives 61% of returns but gets 18% of support spend"

If you cannot write that sentence, you do not have a chart yet. You have two slides, or none. Apply the partner-flip test: titles alone, read in order, must tell the whole story.

### 2. Pick the chart by the message, not by habit
Ask what the message claims, then map it to the chart family in `references/chart-selection.md`:
- comparing items → sorted bar chart (horizontal for long labels)
- showing change over time → line chart; few periods → bar/column
- part-to-whole → stacked bar or small multiples; pie only for 2-3 familiar parts
- distribution → histogram; relationship → scatter

Default to the simplest chart in the family. Never 3D, never exploded pie, never dual unrelated axes on one plot.

### 3. Strip every mark that does not earn its place
The erasure pass, in order:
1. Remove backgrounds, borders, shadows, gradients, decorative icons, 3D effects.
2. Remove gridlines entirely; if the audience must read exact values, keep faint horizontal lines only.
3. Delete the legend; put labels directly on the data (end of lines, on bars).
4. Remove tick marks that do not anchor a reading; reduce axis labels to the fewest needed.
5. Check every remaining mark: does it encode data or carry the message? If not, erase it.

### 4. Annotate so the takeaway is unavoidable
- The title states the conclusion ("South region sales fell 18% after the July price change"), not the topic ("Sales performance").
- One accent color marks the point that matters; everything else stays neutral. Never use color alone to carry meaning: pair it with a label or position.
- Add short annotations on spikes, drops, and turning points ("launch here", "price change"), not a wall of text.
- Put the source in small type at the bottom. Every number on the slide traces to it.

### 5. Frame the slide in the deck's argument
Structure: opening (situation, complication, question, answer), recommendation first, supporting evidence, risks, the ask, appendix. One chart per slide. Every element must pass the "so what" test against the action title. Anything that fails moves to the appendix or gets cut. Before calling it done, run the 10-second test: show the slide, hide it, and ask someone what it was about, what the key number was, and what it recommended.

See `references/before-after.md` for four worked rebuilds and `references/review-checklist.md` for the final audit.

## Output Contract
When asked to produce or review executive chart content, deliver:
1. The action title (one sentence, ≤15 words, key number included).
2. The chart spec: chart type, why this family, sorted/highlighted series, axis and baseline choices.
3. The erasure list: what was removed and why it did not earn its place.
4. The annotation plan: title, accent color target, callouts, source line.
5. The review verdict: pass/fail against `references/review-checklist.md`.

## Operating Rules
1. Message first. No chart exists without an action title.
2. Bars start at zero. If the axis must truncate, mark the break visibly and say so in the note.
3. Check the Lie Factor on any dramatic-looking graphic: graphic % change divided by data % change must stay within 0.95-1.05. Use `bin/lie_factor.py` to compute it.
4. One dimension of data gets one dimension of ink. Never scale pictograms or icons by area/volume to show a scalar.
5. Money over time goes in constant, standardized units. Nominal-only comparisons mislead.
6. Contrast for text ≥ 4.5:1. Never red/green as the only distinction.
7. Source every number. Unsourced data destroys credibility.
8. Never invent data to make an example work. Illustrative data is labeled "illustrative" and stays simple.
9. Bulky detail (raw data, full deck outline, appendix content) goes in references or appendix, never on the slide.