Chart Selection And Encoding

Chart Selection & Encoding — Pick the Right Chart, Then Encode It Honestly

Before any decluttering or colour craft, one decision dominates everything downstream: which chart type does this data and this message actually call for? A beautifully decluttered pie chart is still the wrong chart. This skill routes data-type × message to a concrete chart type, then verifies the encoding is honest (zero baselines, no dual-axis fabrication) and accessible (non-colour cues, contrast, text alternatives). It is the upstream selection decision; the sibling data-visualization owns the downstream craft of drawing the chart you picked.

Use When

• You must choose a chart type for a dataset and a message — and the message is one of: comparison, trend over time, distribution, part-to-whole, relationship/correlation, or geographic.
• You are reviewing or fixing a chart whose type is wrong for its job (pie for ranking, dual-axis combo, 3-D bars, gauge, rainbow heatmap) and need the correct replacement.
• You need to confirm the encoding is honest: zero baseline on length encodings, no secondary y-axis, scale not clipped to manufacture a story.
• You need the accessibility gate for the chosen chart: non-colour cues, contrast floors, text alternative / data table.

Do Not Use When

• The chart type is already settled and you need the craft of drawing it well — decluttering (data-ink ratio, gridline/border removal), the grey-base + one-accent colour rule, preattentive hierarchy, annotation phrasing, or Knaflic's storytelling lessons. Use the sibling data-visualization (it owns the six lessons, the Cleveland–McGill perceptual ranking in references/chart-encoding.md, and the per-chart checklist).
• The question is page-level: which KPIs earn the top row, dashboard zoning, drill-down, cross-filter, or real-time/streaming UX → use dashboard-and-data-product-design.
• It is a generic grid/spacing/layout question (not chart-specific) → use the layout-grid skill; or a colour-ramp construction question → use color-system-and-palette.

Required Inputs

• The data type(s) of the variables to show: quantitative (amount), ordinal (ranked), categorical/nominal (identity), temporal (time), or geographic.
• The message in one sentence — what must the reader see or do? ("Region A outsells B," "revenue is trending up," "these two variables move together," "spending is over budget.")
• The number of series / categories, and whether values share one common scale.
• The medium (live presentation vs circulated PDF vs interactive) and viewport — these change whether you label directly, split, or re-encode.

Workflow

1. Name the message-type first. Classify the message as one of the six: comparison · trend · distribution · part-to-whole · relationship · geographic. The message — not the data shape alone — drives the chart. Per doctrine/design-doctrine.md §2, state the chosen chart and why it fits this message before drawing anything.
2. Route data-type × message → chart type. Load references/chart-fit-decision.md and use its routing table to land on a concrete chart (e.g. ranked comparison of categories → sorted horizontal bar; trend of one continuous series → line; part-to-whole → 100% stacked bar or sorted bar, never pie). The table also lists the banned defaults and their correct replacements.
3. Check the encoding is on the right channel. Confirm the most important comparison sits on the highest-accuracy channel available — position > length > angle > area > intensity > hue. This is the Cleveland–McGill ranking; the full treatment lives in the sibling data-visualization/references/chart-encoding.md §1 — cite it, don't restate it. If your chosen type pushes the key comparison onto angle/area (pie, donut, 3-D), re-route in step 2.
4. Apply axis/scale honesty (Mackinlay expressiveness).
   • Length encodings (bars) require a zero baseline. A non-zero baseline expresses a false ratio (the Fox-News 460%-vs-13% case). No exceptions for bars.
   • Line/dot (position) charts may use a clipped scale if labelled clearly and not over-zoomed — position encodes change truthfully without a zero baseline.
   • Dual-axis ban. Never put two series on a primary + secondary y-axis: it lets you manufacture any crossover by choosing the scales (non-expressive). Replace with two panels sharing one x-axis (small multiples) or direct-label the second series.
5. Annotate the takeaway. Give the chart an action title stating the so-what ("Q3 revenue exceeded target by 12%," not "Q3 Revenue"), title every axis, and call out the one key point directly on the chart. A chart that needs a paragraph to interpret is the wrong chart or is under-annotated.
6. Run the accessibility gate (doctrine/references/wcag-2.2-criteria.md, and the chart a11y floor in data-visualization/references/chart-encoding.md §5):
   • Never colour-alone (WCAG 1.4.1). Back every colour distinction with a non-colour cue — direct label, dash pattern, marker shape, texture, or +/− sign. Use blue (positive) + orange (negative), never red+green, plus a sign/arrow.
   • Contrast: data marks ≥ 3:1 vs background and vs adjacent marks (1.4.11); chart text ≥ 4.5:1 (≥3:1 for ≥24px / ≥18.66px-bold) (1.4.3). Design with APCA, certify with the WCAG ratio.
   • Colour-blind-safe palette (Okabe–Ito / ColorBrewer "colorblind-safe"); ordered data uses a single-hue ramp, never rainbow. Test in a CVD simulator.
   • Text alternative: informative alt/caption (the action title seeds it) + ideally a backing data table (1.1.1); interactive points keyboard-reachable, name/role/value (4.1.2), targets ≥ 24×24px (2.5.8); honour prefers-reduced-motion; usable at 320px / 200% zoom.

Decision Rules

Capability Contract

Read and search are required for data definitions and delivery constraints. Editing is allowed only when chart production is requested. Execution is required for calculations and rendering claims; publication requires separate authority.

Degraded Mode

If required evidence or tooling is unavailable, use the scoped fallback below and mark the result unverified. Without validated data, produce a conditional chart recommendation using a labelled schema. Stop before final encoding when units, grain, or decision question are unknown; mark all unrendered accessibility and scale checks unverified.

Anti-Patterns

• Picking the chart from the data shape alone, ignoring the message — the same table is a bar (comparison) or a line (trend) depending on what you want the reader to see.
• Pie / donut for a comparison or ranking — forces the key comparison onto angle/area (the worst-decoded channels). Replace with a sorted horizontal bar.
• Dual-axis combo chart — two scales chosen to fake a correlation. Split into panels or label directly.
• 3-D bars/pies, gauges, speedometers, rainbow heatmaps — chart-junk that degrades the encoding (doctrine ai-slop-taxonomy.md "dashboard decoration").
• Non-zero bar baseline — visually inflates differences; a truthfulness failure, not a style choice.
• Too many series on one chart ("spaghetti") — re-route to small multiples or emphasise one line; that re-routing is this skill's job, the line-styling craft is the sibling's.
• Meaning carried by colour alone — fails CVD readers and greyscale print.
• Descriptive title with no annotation — "Revenue by Region" instead of the takeaway.

Outputs

• A chart-selection decision: the message-type, the chosen chart type with the data-type × message reason, the named encoding channel for the key comparison, the axis/scale honesty ruling (zero baseline / clipped-OK / dual-axis split), the annotation plan (action title + callout), and the WCAG 2.2 a11y gate result. Hand the chosen type to data-visualization for the drawing craft.

Examples

• examples/chart-selection-cases.md — six worked good/bad selection cases, one per message-type (ranking pie→sorted bar; trend dual-axis→split panels; distribution stacked-bar→histogram/box; part-to-whole pie→100% stacked bar; relationship grouped-bar→scatter; geographic 3-D map→ choropleth), each routed through the data-type × message logic with the honesty and a11y reasoning. Copy the reasoning pattern, never the numbers. Never lorem.

References

• doctrine/design-doctrine.md — the Mission (authored, human-made charts) and the Anti-Slop Charter; the chart type is a stated, defensible choice, never a reflexive default.
• doctrine/references/wcag-2.2-criteria.md — the accessibility floor (contrast, 1.4.1 use of colour, target size, reflow, reduced-motion) every chosen chart must clear.
• doctrine/references/ai-slop-taxonomy.md — the chart-junk / "dashboard decoration" tells (3-D, gauges, rainbow, donut) this skill routes away from.
• Sibling — 12-data-viz-and-dashboards/data-visualization — owns the single-chart craft: the Cleveland–McGill perceptual ranking (references/chart-encoding.md), decluttering, the grey-base + one-accent colour rule, preattentive hierarchy, annotation craft, and Knaflic's six lessons. This skill owns the chart-TYPE selection that comes first; cite that ranking, don't restate it, and hand off the chosen type for drawing.
• Sibling — 12-data-viz-and-dashboards/dashboard-and-data-product-design — owns page-level composition (KPI tiers, zoning, drill-down, cross-filter, real-time UX) above the single chart.
• references/chart-fit-decision.md — the data-type × message → chart-type routing table, the banned-default → replacement map, and the selection decision checklist.