chart-choice

Chart Choice Skill

Chart choice is a sentence-completion exercise: the chart exists to make one stated point legible in three seconds, and the point's grammar picks the form — comparing things (bars), change over time (lines), parts of a whole (stacked/100% bars — pie only under five slices), distribution (histogram), relationship (scatter). Most bad charts are either the wrong form for the point, or the right form dressed dishonestly: truncated bar axes, dual axes engineered for drama, 3D anything. The skill picks the form, then runs the honesty pass.

What This Skill Produces

• The chart verdict — the form, chosen from the point's grammar, with the reasoning stated
• The honesty pass — axes, baselines, sorting, and the dual-axis/3D vetoes applied
• The labeling spec — title-as-the-point, direct series labels, the source line — so the chart survives being screenshot out of context
• The alternative — when the honest answer is "this wants a table, not a chart"

Required Inputs

Ask for these if not provided:

• The point, as a sentence: "show the data" isn't chartable; "East region drove the Q2 recovery" is — the sentence picks the chart, and extracting it is half the skill
• The data's shape — categories × measures × time; how many series (five lines are a chart; twelve are spaghetti — the shape forces choices)
• The venue — a live deck (bolder, fewer labels), a doc (denser is fine), a dashboard (self-serve labeling); and whether it will be screenshot-forwarded (assume yes)

Framework: The Choice and Honesty Rules

1. The grammar mapping: compare items → horizontal bars, sorted by value (not alphabet) · trend → line, one-to-five series · composition → stacked bar (100% when shares matter more than totals; pie only ≤5 slices and only when "half" or "quarter" is the point) · distribution → histogram · relationship → scatter. When the point contains two of these, that's two charts.
2. Bar axes start at zero — line axes may not: bars encode value as length, so truncation lies structurally; lines encode change as slope, so a zoomed line axis is legitimate when labeled. This asymmetry resolves most axis arguments.
3. The dual-axis veto (nearly always): two measures on two axes lets the author choose the visual correlation by scaling — the honest alternatives are two stacked panels or an indexed chart (both series = 100 at start). Exceptions are rare and carry labels.
4. Sort for the point, declutter to the point: categorical bars sorted descending unless order is inherent (months) · gridlines light, legend replaced by direct end-of-line labels when few series, decimals only to meaningful precision · 3D never — depth encodes nothing and distorts everything.
5. The title is the point, not the topic: "Revenue by region" is a topic; "East drove the Q2 recovery" is the point — titled so, with the source-and-date line at the foot, the chart survives forwarding without its author. The final test: a stranger sees it for three seconds — do they say your sentence back?

Output Format

Chart: [the point, as the sentence]

The Verdict

[Form + the grammar reasoning · series count check · the two-charts split if the point was compound]

The Honesty Pass

[Axis baselines (bar-zero / line-labeled) · dual-axis check → panels/indexed if triggered · sort order · the 3D/decoration vetoes]

The Labeling Spec

[Title = the sentence · direct labels vs legend · precision · source + as-of line]

The Three-Second Test

[What a stranger reads off it cold — and the fix if that isn't the point]

Quality Checks

• [ ] The point exists as a sentence before the form was chosen
• [ ] Bar charts baseline at zero; zoomed line axes carry their label
• [ ] No dual axes without the stated rare-exception justification
• [ ] The title states the point; the source line exists
• [ ] Charts with 6+ series were split, aggregated, or turned into a table

Anti-Patterns

• [ ] Do not chart a topic — no point-sentence, no chart; a table shows data that hasn't decided its point
• [ ] Do not truncate bar axes for drama — that's lying in geometry
• [ ] Do not engineer dual-axis correlations — panels or indexing tell it straight
• [ ] Do not pie past five slices — comparison by angle fails exactly when slices multiply
• [ ] Do not decorate — every ink drop that isn't data competes with the three seconds the chart gets