agentleFS
jpoindexterjpoindexter/tufte-skills/skills/tufte-graphical-excellence/SKILL.md

tufte-graphical-excellence

Skill1 starsChanged 4 months agoMITView on GitHub

Use when designing a new data graphic, grading or reviewing an existing chart, or deciding whether data deserves a graphic at all — especially when a chart looks polished but feels thin, shows a single variable, forces a slow read, or when summary statistics are being offered in place of a plot of the actual data.

The file

We split this skill into the rules it gives an agent

2 rules. About 0 are common, 2 are distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈2

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Treat anything outside roughly 0.95–1.05 as distortion to fix.

    line 209
  2. Distinctive:

    Variable trend → DO: use a time-series for richly varying data (a year of daily temperature swings).

    line 217

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Graphical Excellence

Overview

Graphical excellence is Tufte's unified standard for a data graphic, set out in Chapter 1 of The Visual Display of Quantitative Information. A graphic's quality is not how striking it looks but how many true ideas it delivers per unit of viewer time, ink, and space. Excellence is simultaneously a matter of substance, statistics, and design; it is nearly always multivariate; and it is conditional on telling the truth about the data. This skill gives an AI agent the criteria, the optimization formula, the exemplary forms, and the failure modes needed to design or grade any chart.

"As for a picture, if it isn't worth a thousand words, the hell with it." — Ad Reinhardt, quoted in Visual Explanations, p. 119

Tufte calls that possibly the best sentence ever written about the point of images for information design — the excellence standard compressed to a threshold test.

How excellence is distributed (SWFE, p. 47). Seeing With Fresh Eyes adds a sobering statistical frame: excellence is scarce, its distribution lognormal and long-tailed — most work clusters near the bottom and a thin tail carries the rest — but once reached it is forever knowledge, not a fashion that expires. The practical corollary Tufte draws: because analytical thinking is fundamentally about the relationship between evidence and conclusion, credibility can sometimes be judged without expertise in the subject matter itself.


§1. The Five Criteria (VDQI p. 51)

Tufte's closing summary of the chapter states five criteria. They are constraints, not preferences — failing any one disqualifies a graphic regardless of how well it meets the others. Paraphrased:

  1. Substance + statistics + design. Excellence is the well-designed presentation of interesting data — all three dimensions at once, never one compensating for another.
  2. Clarity, precision, efficiency. Complex ideas should arrive communicated clearly, accurately, and without wasted effort.
  3. Maximum ideas, minimum cost. Give the viewer the most ideas in the least time, with the least ink, in the smallest space (the formula in §2).
  4. Multivariate by default. One short, attributed quote captures it exactly:

    "Graphical excellence is nearly always multivariate." — Tufte, The Visual Display of Quantitative Information

  5. Truth. Excellence requires telling the truth about the data; distortion voids it (§10).

A graphic can be beautiful and still fail by being univariate, or honest and still fail by taking ten minutes to surrender three numbers.

The longer canon — nine "Graphical displays should…" principles (p. 13)

Before the p. 51 five, the chapter opens with a nine-item list of what graphical displays should do (paraphrased): show the data; steer the viewer to the substance rather than to the method or the design; avoid distorting what the data say; present many numbers in a small space; make large datasets coherent; nudge the eye toward comparing different pieces of the data; reveal the data at several levels, from broad overview to fine structure; serve a reasonably clear purpose — description, exploration, tabulation, or decoration; and mesh tightly with the statistical and verbal descriptions of the same dataset. Treat p. 13 as the working checklist and p. 51 as the closing verdict.


§2. The Formula: Ideas / Time / Ink / Space

Criterion 3 is a concrete four-axis optimization target.

AxisDirectionWhat it demands
IdeasMaximizeMore variables, more comparisons, more structure surfaced
TimeMinimizeGrasped fast — little or no decoding required
InkMinimizeEvery mark earns its place; non-data ink is erased
SpaceMinimizeHigh density — pack more data per square centimeter

The four axes pull against each other; the craft is resolving the tension without sacrificing any axis. Many ideas that need long study fail the time axis. A fast read that shows one variable fails the ideas axis.

Benchmark exemplar — Minard, 1869. Charles Joseph Minard's flow map of Napoleon's 1812 Russian campaign hits all four axes at once, encoding six variables in one frame (detailed in §7). Étienne-Jules Marey praised it as seeming to defy the historian's pen with its blunt eloquence. Tufte's verdict:

"It may well be the best statistical graphic ever drawn." — Tufte, The Visual Display of Quantitative Information


§3. The Three Required Dimensions: Substance, Statistics, Design

Criterion 1 names three dimensions that must all be present. The diagnostic value is in the failure mode when one is missing.

DimensionFailure mode if absentConcrete check
SubstanceWell-drawn graphic of data too thin to deserve oneIs this worth a graphic, or would one or two numbers / a table do?
StatisticsInteresting data whose structure is misrepresented or flattenedAre variability, comparison, and multivariate structure preserved?
DesignSound data whose presentation hides rather than revealsCan a cold viewer read the structure in seconds?

Substance failure is the most common and least noticed: a polished chart of a single smooth trend that two endpoints would summarize.


§4. A Short History (graphics are recent)

A point easy to miss: statistical graphics are a young invention — roughly two centuries old — even though maps and pictures are ancient. Tufte traces the lineage so the reader treats the form as a designed technology, not a given.

OriginContribution
Oldest known time-seriesA tenth- or eleventh-century manuscript graph charting the inclinations of the planetary orbits as a function of time (p. 28) — an isolated precursor that left no tradition.
Michael Florent van Langren (1644)Visual Explanations (p. 15) nominates his longitude chart as the first statistical graphic: twelve rival estimates of one longitude difference laid out along a single line. The true value is 16°30′ — and every one of the twelve overstates it. A first graphic that is also a first demonstration of systematic bias made visible by putting the estimates side by side.
J. H. Lambert (18th c.)Early systematic use of line graphs / time-series to show measured relationships; VE (pp. 15–16) credits him with quantifying the plane itself — putting measured scales on both axes.
William Playfair (1786, 1801)Inventor of most modern statistical graphic forms — the time-series line graph and the bar chart in the Commercial and Political Atlas (1786), the pie and circle charts in the Statistical Breviary (1801).

Playfair also argued the case for graphics over tables: numbers studied in a table leave only a faint impression, like a figure drawn in sand, while a picture imprints and is retained. The takeaway for an agent: the conventions you reach for reflexively were deliberately designed within recorded history, and can be redesigned.


§5. Graphics vs. Tables — When Each Wins

Canonical owner: tufte-tables-and-text-tables carries the full format ladder (sentence → text-table → table → semi-graphic → graphic), the twenty-number rule with its perceptual grounding, and the supertable grammar. This section keeps only the graphic/table decision as it bears on grading a graphic.

Excellence sometimes means not drawing a graphic. Tufte is explicit that small, sparse, or non-comparative data belong in tables.

SituationUse graphicUse table
Large N with real structure and variabilityYesNo
Multivariate relationships to compareYesNo
Bivariate or higher relational dataYes (scatterplot)No
Small N, highly labeled, non-comparativeNoYes
Single linear trend; 1–2 numbers sufficeNoYes (or just the numbers)

Numeric threshold (p. 56). Tables usually outperform graphics for datasets of roughly twenty numbers or fewer — the book's own cutoff for where the graphic form starts to earn its keep.

Anscombe's quartet (pp. 13–14) — why "just report the regression" fails. Four small datasets share the same means, fitted regression line, and correlation to two decimal places, yet plotted they are four utterly different structures: a well-behaved linear cloud, a smooth curve, a tight line wrecked by one outlier, and a vertical stack with a single leverage point. Identical summary statistics, completely different data behavior. Graphing is an act of analysis, not decoration — the graphic reveals what the numbers conceal — so the quartet is the standing rebuttal whenever a table of summaries is offered in place of a plot.

Density threshold (the case for the graphic). The New York City weather summary for 1980 packs 1,888 numbers into one chart — daily highs and lows against the long-run normal, plus precipitation and humidity — and lets the viewer compare any part to any other and read a year as a story. No table does that. Use a graphic when it organizes a large collection of numbers, supports comparison across parts, and tells a story; otherwise prefer the table.


§6. The Four Exemplary Graphic Forms

VDQI Chapter 1 surveys four major forms and tests each against the excellence criteria. (Small multiples is a Chapter-8 technique, not one of these four.)

1. Data maps

The densest form. The U.S. cancer-mortality atlas places an age-adjusted rate for each of 3,056 counties on a single map — roughly 21,000 numbers per map once rate, location, and county size and shape are counted (p. 17) — with separate maps per cancer site and sex. A volume of data no other display can carry in the space. Data maps read at several scales at once: continental gradient, regional cluster, single county.

  • John Snow's 1854 cholera map plotted deaths as dots and pumps as marks; the cluster around the Broad Street pump was visible instantly, where calculation would have needed luck and labor to find the same result. Graphical analysis can testify about data far more efficiently than computation.
  • Caveat (anticipates the blot-map failure, §9): shaded-area maps weight by geographic area, not population — sparse western counties loom large, dense eastern ones recede.

2. Time-series

The dominant published form: more than three-quarters of graphics in major newspapers and magazines (surveyed 1974–1980) were time-series. The natural left-to-right ordering of the time axis gives an interpretive ease no other arrangement matches. Failure mode: a smooth monotonic change that one or two numbers capture does not justify the form — reserve time-series for genuinely variable, complex material.

3. Narrative graphics of space and time

The fourth form combines geography and chronology in one frame — the category Minard's Napoleon map belongs to, alongside Marey's graphical Paris–Lyon train schedule (time on one axis, distance/stations on the other, each train a diagonal whose slope is its speed). These narrate movement through space over time, doing what neither a pure map nor a pure time-series can. This is the natural home of the multivariate masterpiece in §7.

4. Relational graphics (scatterplot)

About 40% of graphics in scientific publications are relational. Tufte rates the scatterplot the strongest design of all:

"the greatest of all graphical designs." — Tufte, The Visual Display of Quantitative Information

It links two or more variables and pushes the viewer to weigh a possible causal relationship. Example: per-capita cigarette consumption (around 1930) plotted against male lung-cancer death rates (around 1950) across eleven countries — the strong positive relationship is obvious at a glance and inert in the equivalent table.


§7. Multivariate as the Default Standard

Criterion 4 is a grading rule, not an aspiration: showing one variable where two or more are structurally linked is a failure.

VariablesTypical formVerdict
1Bar / pie chartUsually better as a table or a single number
2ScatterplotMinimum threshold for the graphic form
3–4Layered time-series, annotated mapStrong candidate
5–6Space-time narrative (Minard)The exemplary form
10⁵+Data mapGraphics have no competitor

Minard's six variables, woven into one band:

  1. Army size — width of the band
  2. Longitude — horizontal position
  3. Latitude — vertical position
  4. Direction of march — color (advance vs. retreat)
  5. Temperature during the retreat — a linked lower line graph
  6. Time — dates tied to position along the retreat

The narrative payload of those six variables (pp. 40–41): 422,000 men cross into Russia, 100,000 reach Moscow, 10,000 return. The reader absorbs a four-to-five-dimensional account of the campaign without ever feeling the complexity of what is being read. That seamlessness is the goal: multivariate depth that does not announce its own technique.


§8. What Graphical Excellence Is Not

Opening Chapter 2 (p. 53), Tufte names two wrong assumptions that produced the graphically barren years from 1930 to 1970:

  • Wrong assumption 1 — graphics must be "livened up" with decoration and exaggeration or the audience drifts.
  • Wrong assumption 2 — the analyst's main job is detecting and denouncing lies. (Integrity is necessary, but it is a floor, not the point.)

The audience premise (EI p. 34). Both wrong assumptions rest on contempt for the reader. Tufte's corrective is to design on the opposite premise: readers are alert and caring, possibly busy and eager to get on with it, but not stupid.

"Clarity and simplicity are completely opposite simple-mindedness." — Tufte, Envisioning Information, p. 34

Excellence therefore aims at invisibility — a design so good it disappears, leaving attention on the data rather than on the data-container (EI p. 33), with the standard set by high-quality maps rather than by posters (EI p. 35).

Not excellenceWhy it fails
Decoration / ornamentNon-data ink steals attention from data ink
"Pretty" independent of contentAesthetics unmoored from data add noise
One dramatic summaryA good graphic reveals more than any single statistic could
Univariate displayOne variable rarely justifies the medium; a table beats it
Mere chronologyPassage of time is not causal explanation
DistortionTruth is a precondition, not a bonus

§9. Named Failure Modes

Failure modeDescriptionIllustration
Substance failureData too thin to deserve a graphicA single smooth trend that two numbers summarize
Decoration failureNon-data ink dominates the surfaceBars with photos behind them; 3-D shading on flat data
Univariate failureOne variable shown where more are neededA lone time-series with nothing to compare against
DistortionVisual magnitude ≠ numerical magnitudeTruncated/hidden baseline; perspective inflating differences
Chronological confusionTime-series read as causal when time isn't the causeA trend actually driven by an unshown third variable
Table-as-graphicSmall, labeled, non-comparative data forced into a chartSparse data that a table would serve better
Blot-map problemArea shading weights by geography, not by population/countCancer map where empty large counties dominate dense small ones

§10. The Truth Requirement (bridge to integrity)

Truth is criterion 5 and a hard gate: a graphic that misrepresents its data cannot be excellent no matter how multivariate or ink-efficient it is. The disappearing baseline, the perspective trick that inflates differences, the Nobel-prizes chart that counts prizes in decade intervals but ends on a 4-year interval — manufacturing a fake final decline, which Tufte flatly calls a big lie (p. 60) — each fails at the foundation, and no sophistication elsewhere rescues it.

The quantitative measure (developed in Chapter 2, Graphical Integrity) is the Lie Factor. It is restated here only as the gate on criterion 5 — tufte-graphical-integrity §1 is its canonical treatment, with the worked computations, the distortion catalog, and the audit protocol:

Lie Factor = (size of effect shown in the graphic) / (size of effect in the data)
  • = 1.0 — truthful.
  • > 1.0 — exaggerated (the usual case).
  • < 1.0 — understated.
  • Treat anything outside roughly 0.95–1.05 as distortion to fix.

§11. Application: Do / Don't and Checklist

Named do / don't pairs (concrete):

  • Relationships → DO plot two linked variables as a scatterplot so the correlation is visible at a glance. DON'T bury the same paired numbers in a table where the relationship stays hidden.
  • Variable trend → DO use a time-series for richly varying data (a year of daily temperature swings). DON'T spend a time-series on a smooth monotonic trend that two endpoints capture.
  • Spatial pattern → DO map events over geography (cholera deaths over streets) to expose the cluster. DON'T rely on summary statistics and hope the spatial structure falls out — it usually won't.
  • Linked variables → DO weave causally-related variables into one frame (Minard). DON'T scatter them across separate univariate charts that force the reader to reintegrate them mentally.
  • Thin data → DO ship a table or the bare numbers. DON'T dress one or two values up as a chart.

Pre-publish checklist:

  • Is the data interesting and complex enough to deserve a graphic, or is a table/number better? (substance)
  • Are at least two variables shown? If one, is the graphic form justified? (multivariate)
  • Can a cold viewer read the structure in ~10–30 seconds? (time)
  • Is every mark data-bearing — any non-data ink to erase? (ink)
  • Is density high — many data points per square centimeter? (space)
  • Does visual magnitude match numerical magnitude — Lie Factor ≈ 1.0? (truth)
  • Does it reveal something a single summary statistic would hide? (statistics)
  • Did I pick the right form for the job — map, time-series, space-time narrative, or scatterplot? (§6)

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.