Data Visualization Guide

A decision framework for turning data into graphics that communicate. Optimized for business and analytics work — reports, dashboards, slides, and the exec-facing "what should we do?" chart — but the principles are tool-agnostic and apply in Tableau, Python, R, JavaScript, or a spreadsheet.

The guidance distills the standard literature on the subject. See references/BOOKS.md for the fifteen books behind it (Knaflic, Wilke, Munzner, Cairo ×2, Dykes, Berinato, Camões, Setlur & Cogley, Meirelles, Krum, Duarte, Dougherty & Ilyankou, Khan, Kirk) and which file draws on each.

These are principles of human perception and communication, not chart-tool trends. They hold regardless of library or era.

When to use

• Creating a chart, graph, plot, map, or infographic from data
• Reviewing or critiquing an existing visualization
• Choosing which chart type fits a question and a dataset
• Picking or fixing colors, axes, scales, labels, or annotations
• Designing a dashboard or a single-screen monitoring view
• Turning analysis into a presentation, report, or data story that drives a decision

When NOT to use

• Building data pipelines, ETL, or wrangling raw data into shape
• Statistical modeling, hypothesis testing, or choosing analysis methods
• Selecting a BI platform or visualization vendor
• General interface layout, navigation, or interaction design — use ux-design-principles

The core idea

A visualization is an act of communication, not decoration. Before drawing anything, you owe three answers (Munzner's what / why / how):

1. Who is the audience and what decision or insight is this for? ("Why")
2. What is the shape of the data? (quantities, categories, time, parts, relationships, geography) ("What")
3. What encoding shows that relationship most accurately to the human eye? ("How")

Get 1 and 2 right and the chart type often picks itself. Get the encoding right and the reader sees the point without effort. Everything else in this skill serves those three questions.

Explain-the-why runs throughout. Each rule is paired with the reason behind it, so you can apply it to cases this guide never anticipated. A rule you can't justify is one you'll misapply.

Workflow

Use this loop when producing a visualization. Paste it and tick boxes for any non-trivial chart.

[ ] 1. State the question/decision in one sentence. Name the audience.
[ ] 2. Identify the data shape (see references/chart-selection.md).
[ ] 3. Pick the chart family from intent + data shape.
[ ] 4. Choose the encoding for accuracy (references/perception-and-encoding.md).
[ ] 5. Declutter: remove anything that isn't data or doesn't aid reading (references/decluttering-and-integrity.md).
[ ] 6. Direct attention: title states the takeaway; color/annotation highlight the point.
[ ] 7. Check integrity: axes, scales, proportions honest? (references/decluttering-and-integrity.md).
[ ] 8. Check accessibility: colorblind-safe, labeled, sufficient contrast (references/color-and-accessibility.md).
[ ] 9. Re-read as the audience. Is the message obvious in 5 seconds?

When reviewing a visualization, run the loop backwards as a critique: does the title carry a message? Could a different encoding be read more accurately? Is any ink not earning its place? Is anything misleading? Adopt a reviewer persona — you are a data-viz editor whose job is to make the single most important point land truthfully.

Reference map (load on demand)

Read the file that matches the task — don't load all of them at once.

Fast defaults

When you have no time to read further, these defaults are right far more often than not. Each links to the file that explains why and when to break it.

• Comparing values across categories → horizontal bar chart, sorted by value (not alphabetically). Bars encode magnitude as length/position, the most accurately read channel.
• Change over time → line chart. Time on x, one line per series; label lines directly instead of a legend.
• Part-to-whole → stacked bar or just a bar chart of the parts. Avoid pie charts beyond ~2–3 slices; angle and area are read poorly (see perception file).
• Relationship between two measures → scatter plot. Add a trend line only if you mean it.
• Distribution → histogram (one variable) or box/violin (comparing groups).
• Bar charts start at zero, always — bar length is the value, so a clipped baseline lies. Line charts may use a non-zero baseline because they encode position/slope, not length.
• One message per chart. If it needs two takeaways, it's two charts.
• The title is the takeaway, written as a sentence ("Revenue fell 12% after the May price change"), not a label ("Revenue over time").
• Default to a colorblind-safe, perceptually-uniform palette (e.g. viridis for sequential). Never encode meaning by color alone.
• Grey by default, one accent color for the thing that matters; direct-label series instead of a legend; round numbers to what the decision needs ($1.2M, not $1,234,567); faint or no gridlines.

For a worked example of each situation above — rendered chart plus the exact code — see references/chart-gallery.md.

Known gotchas

Specific traps that look fine until someone is misled. Details and fixes in the linked files.

• Truncated bar axis. A bar chart not starting at zero exaggerates differences. (Line charts are the opposite — forcing zero can flatten a meaningful trend.) → decluttering-and-integrity.md
• Dual y-axes. Two series on two scales invites readers to infer a correlation you manufactured by choosing the scales. Prefer indexed series or two stacked panels. → decluttering-and-integrity.md
• Rainbow / jet colormaps create false boundaries and aren't perceptually uniform; the "bright band" looks like a feature in the data that isn't there. Use viridis/cividis. → color-and-accessibility.md
• Encoding quantity by area or radius. Doubling a bubble's radius quadruples its area; readers under-read area by a known factor. Scale by area, and prefer length where you can. → perception-and-encoding.md
• Pie charts with many slices, or 3-D anything. Angle, area, and volume sit at the bottom of the accuracy ranking; 3-D adds occlusion and perspective distortion for zero information. → perception-and-encoding.md
• Legends far from the data. Forcing the eye to ping-pong between a legend and the plot adds cognitive load; label series directly. → storytelling.md
• Overplotting. Thousands of opaque points become a blob. Use transparency, binning (hexbin), or sampling. → chart-selection.md
• Cherry-picked ranges and unlabeled log scales. Both are integrity failures even when each data point is accurate. → decluttering-and-integrity.md

A worked example (good vs. bad)

Task: "Show the board how our five regions performed against quota this quarter."

• Bad: A 3-D pie chart, one slice per region, rainbow colors, legend on the side, title "Regional Performance." The reader can't compare slice sizes (angle/area), the 3-D tilt distorts them further, and the title carries no message.
• Good: A horizontal bar chart, one bar per region showing % of quota, sorted descending, a reference line at 100%, the one region that missed quota colored to stand out and the rest in grey, every bar directly labeled, baseline at zero. Title: "Four of five regions beat quota; West missed by 8%." The encoding (length on a common scale) is the most accurate available, color directs attention to the single exception, and the title states the decision-relevant fact.

This is the whole method in miniature: right question → right data shape → most-accurate encoding → declutter → direct attention → honest axis → takeaway title.