agentleFS
jpoindexterjpoindexter/tufte-skills/skills/tufte-typography-for-data/SKILL.md

tufte-typography-for-data

Skill1 starsChanged 4 months agoMITView on GitHub

Use when setting type that carries data — chart labels, axis titles, table headers, legends, captions, schedules, maps, dashboards, or numeric columns; when a legend, key, or letter-coded caption is being added; when linebreaks, alignment, or column width are being decided by a grid or style sheet rather than by content; when choosing a typeface or weight for small or dense text; or when a display's aspect ratio is being set by a default.

The file

We split this skill into the rules it gives an agent

6 rules. About 0 are common, 5 are distinctive, and 1 is very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈5

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 1

    Very specific

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

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Every number in this table is Tufte's own research.

    line 79
  2. Distinctive:

    Never a disconnected legend box when direct placement is possible.

    line 230
  3. Very specific:

    Design rule: when evidence is contested, give the rebuttal a distinct color and put it next to the claim it rebuts, not in a separate …

    line 257
  4. Distinctive:

    Annotation density test: remove every label. If the display is still fully self-explanatory, the labels were redundant.

    line 345
  5. Distinctive:

    F13 — Uniform paragraph styling: Every paragraph typographically identical, defeating retrieval.

    line 396
  6. Distinctive:

    All passages paraphrased; quoted lines limited to single attributed sentences with page citations.

    line 434

≈ “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

Typography for Data Display

Overview

Typography in an analytical display has one job: serve the content, never the production process. Tufte's governing distinction (Seeing With Fresh Eyes, Chapter 2, printed pp. 49–63) is between content-responsive typography — where space, linebreaks, and label placement are driven by the data and its structure — and content-indifferent or content-hostile typography, where generic grids, squared-off margins, and legend boxes subordinate meaning to convenience of layout.

The standard is borrowed from typographer Robert Bringhurst, and it appears in Chapter 1, p. 19, as the epigraph over the Cleveland Clinic thought-bubble spread — not in Chapter 2:

"Typography exists to honor content." — Robert Bringhurst, The Elements of Typographic Style (1992), p. 17, quoted in Tufte, Seeing With Fresh Eyes, p. 19

Two distinct claims that are routinely fused — keep them apart:

ClaimFigurePageWhat it is about
Printed text has used grids indifferent or hostile to meaning1500 yearsp. 17Control — publishers, text-grids, and style-sheets own the space
Space was costly, so words ran edge to edgethousands of yearsp. 53Cost — and on screens the marginal cost of space is now zero

The 1,500-year figure belongs to the control argument (the Great Disjunction), not to the cost of paper.

Companion skills: tufte-visual-thinking (the model-criticism frame this sits inside), tufte-annotation-as-thinking (Ch. 5), tufte-data-maps (Swiss topo maps in full), tufte-parallelism.


§1. Four Modes, Not Two

Tufte's scale runs further in both directions than a responsive/hostile pair suggests. Content-possessed typography (p. 51) is the far end, and it is a legitimate category, not an error.

ModeMechanismEffectSource
Content-indifferentWordspaces stretched to square off the right marginSyntactic structure lost; word saladp. 17
Content-hostileLinebreaks chop phrases mid-clause to fill a columnMeaning actively suppressedp. 17
Content-responsiveLinebreaks follow commas, semicolons, full stops; spacing exposes parallel structureMeaning amplified; syntactic units visiblepp. 17, 52
Content-possessedThe typographic form becomes the thing describedApollinaire's calligramme sets "Il pleut" as five slanting vertical columns of single letters, so the text physically rainsp. 51

Boundary for data work: possession is legitimate when the form is the datum — rain, terrain, stroke order, a descent, a rise. It is decoration the moment the form asserts something the data does not. (Related: the 2-D typography of Barr's cubism chart and Apollinaire's Calligrammes permits multiple simultaneous reading paths — BE p. 66; see tufte-narrative-and-sequence.)


§2. Linebreaks That Honor Syntax

Words and sentences interact with the spaces and the other words around them. Text can survive content-indifferent spacing, but survival is not the goal.

Do / Don't — linebreaks in a list sentence:

PracticeVerdictWhy
Cram a 6-item list into one justified paragraphContent-hostilePunctuation alone must carry the structure; reader cannot scan or compare items
Break each list element onto its own lineContent-responsiveThe structure already in the prose becomes visible; items align for up/down comparison

The Ioannidis example (pp. 49 and 52; previewed p. 17). John Ioannidis's claim that most published research findings are false is stated as a six-condition list: smaller studies; smaller effect sizes; a greater number and lesser preselection of tested relationships; greater flexibility in designs, definitions, outcomes, and analytical modes; greater financial and other interest and prejudice; more teams chasing statistical significance. Set as one justified block, five semicolons carry all six elements and readers cannot pull them apart. Give each condition its own line and the list becomes scannable — nothing added, only the existing punctuation honored.

The Somerset Maugham example (pp. 48 and 52). A passage from "The Book Bag" catalogues book types across roughly fourteen clauses. Run together, the categories vanish; broken at the existing commas, semicolons, and stops, the parallel syntax (illness, travel, expedition) becomes visible. The punctuation already marks where the breaks belong.

Chapter geography: Chapter 2 opens on p. 49. The Maugham prelude spread sits on the facing p. 48, before the chapter opener. The KMP code (p. 14) and the poetry-line critics (p. 15) are Chapter 1 material, not Chapter 2.


§3. The Darwin Calculation — Scale of the Problem (p. 50)

QuantityValue
Chapters14
Topics announced in the Table of Contents133
Times each topic is printed (TOC + chapter head)2
Total words of topic lists1,962
Cost to re-set them content-responsively+2 to 3 pages in a 500-page book
Readers since 1859~3 million serious + millions of skimmers
Cumulative reader-impediments imposed~2,000,000,000 to 5,000,000,000

A content-hostile linebreak is never neutral — at scale it is either imposing friction or removing it.

Internal inconsistency in the book, worth knowing: the p. 50 headline says content-responsive typography eliminates trillions (10¹²) of impediments, while the body text on the same page computes 2–5 billion. The body figure is the one derived from the arithmetic and the one used above.

Every number in this table is Tufte's own research; keep the p. 50 citation attached wherever they travel.


§4. Zero-Cost Space and Alignment as Signal (p. 53)

For millennia, space on clay, papyrus, vellum, and paper was costly, so words ran edge to edge. On a display screen the marginal cost of an additional space is zero. Continuing to impose content-indifferent grids on screens honors a constraint that no longer exists.

Alignment modeSignal valueBest use
Justified (both margins squared)None — wordspaces stretched to fillDense reference text; truly cost-constrained media
Ragged-right (flush left)Linebreak position can signal meaningGeneral analytical prose
Central-axis (centered)Both margins active; readers and speakers do not have to hunt for the left margin, and are less likely to skip a line, because the next line's full length and pace are visible in advancePoetry, announcements, data briefings
Ragged-left (flush right)Left space identifies the speakerDialogue in novels and scripts, transcripts

§5. Authors Control the Breaks — Poetry, Code, and the Math Style Sheet

Poetry as the model (p. 15). Critics treat the line — not meter or rhyme — as the feature that makes poetry read as poetry; the line is the most sensitive register of breath and pacing and carries meaning vertically as well as horizontally. Where a line breaks is information, not decoration.

Code as the model (p. 14). A clean Knuth–Morris–Pratt string-matching implementation is roughly 37 lines and about three-quarters empty space. For the reader those spaces create fluency; for the machine they mean nothing and are discarded at compile time. Minify it and the program runs identically and becomes unreadable. Whitespace in an analytical display is for the human, not the medium.

The math style sheet (p. 17) — the strongest external authority in the book. A formula too long for the page width must be broken across lines, and the rules are precise:

  1. Break at the end of a natural phrase.
  2. In a heavily bracketed formula, break at the end of a major bracket, never at an inner symbol.
  3. The rationale is explicitly musical: natural phrasing, as in music or speech, is what makes the break intelligible between writer and reader.
  4. The decision must not be left to the compositor. The author, who knows the content, indicates the point of fracture in advance.

Direct transfer to data work: long axis titles, wrapped table headers, multi-line chart labels, and tooltip copy should all carry author-specified break points, not renderer-chosen ones. If your charting library wraps by width alone, you have handed the compositor the decision.


§6. Regulating Reading Speed (p. 57)

Typography controls not only clarity but pace.

  • Minify hyphens; never minify dashes. A hyphen carries little meaning and should be reduced where possible: e-mail → email. A dash must survive, because a dash's pause can be as strong as a poetic line break. Tufte demonstrates by re-setting a Kingsley Amis sentence so each dash-bounded clause lands on its own line.
  • The suspended sentence. Amis on Elmore Leonard: the present participle suspends the English sentence, opening a lag in time the reader slides through. The typographic corollary (p. 67) is explicit — spacing should maximize explanatory depth, not speed-reading; a graphical sentence there uses 11 line breaks, 3 stacklists, and 40 content-spaced words.

Use for data display: deliberate spacing is a legitimate way to force a slower, comparing read of a dense claim. Speed is not the objective function.


§7. Idiosyncratic Paragraphs and Memory Retrieval (p. 59)

Text-only paragraphs are typographically identical — same face, same leading, same measure, stacked into columns. Tufte's mechanism for why that fails is specific and worth stating precisely: idiosyncratic paragraphs aid memory and retrieval by uniquely activating the relevant neural substrates for retaining visual memories. The reader's failure is equally precise: being unable to find and re-read a specific string of words in a paragraph they have already read.

Tufte deliberately gave nearly every paragraph in the book a unique appearance for exactly this reason.

Do / Don't — paragraph variety:

PracticeVerdictWhy
All body paragraphs same face, leading, measureContent-hostileNo visual anchors; the reader cannot re-find a passage
Tables formatted identically to block proseContent-hostileEvidence types become visually indistinguishable
Tables, code, callouts, quotations in distinct treatmentsContent-responsiveEach evidence type leaves a distinct visual memory
Deliberate spacing between conceptual groupsContent-responsiveGrouping is visible; transitions are marked

Mode-switching mid-sentence, and stacklists inside prose (p. 59). The unit of typographic variation is the paragraph, and variation can happen without breaking the surrounding text:

  • Galileo changed information mode mid-sentence in his 1612 manuscript on Saturn's rings, drawing the observed shapes inline; his publisher then integrated text and images in print.
  • A Lucky Jim passage is re-set so three parallel interrogatives lift out of the justified block into a centered three-line stacklist, leaving the prose around them intact.

§8. The White Interval — Grounding Outside Western Typography (p. 54)

  • György Kepes: Chinese and Japanese painters leave large paths of the picture surface empty and divide it at unequal intervals, which forces eye movements of varying velocity and unifies the surface through maximum variation.
  • Characters are written inside imaginary squares, with the blank areas given as much design attention as the strokes.
  • Tufte's paired example: an 1808 Hokuba woodblock in which calligraphy drifts through the perspective space above the bowls, so the receding planes accent the drifting text.

Transfer: unequal intervals do more work than equal ones. A dashboard on a uniform 8pt rhythm reads as one undifferentiated field; varying the interval by conceptual grouping does the separation that borders would otherwise be asked to do.


§9. LittleDataGraphics and Data Paragraphs (pp. 60–61)

The LittleDataGraphics critique (p. 60). Tufte's term names the family of displays — pie charts, small bar charts — that encode a tiny data set into areas, angles, and colors, forcing the viewer to mentally reverse the encoding. The code is local to that one graphic and repays no learning.

Display typeEncodingDecoding requiredVerdict
Pie chartData → area / angleMental reconversion to proportionContent-hostile: one-off local code, zero transfer value
Bar chart (small n)Data → heightMental reconversion to magnitudeContent-hostile: one-off local code, zero transfer value
Data tableData → numberNone — the number is the dataContent-responsive
Data paragraphData → words + numbers + small imageNone — the prose is the dataContent-responsive

"Data visualizations are at their best when there is so much data that the only way to see it . . . is to see it." — Tufte, Seeing With Fresh Eyes, p. 60

Three worked data paragraphs — keep them distinct:

ExampleContentsp.
NWS KSFO point briefing18 measurements (US and metric), 28 words, 4 links; serves the casual reader via a sunny/cloudy icon and the analyst via the 5th significant digit of barometric pressure, with linked 3-day history and hourly forecast60
Egyptian vultureOne paragraph combining five information sources spanning 4,000 years — an ancient line-and-tone illustration, three sentences of drawing instructions, six stop-action stroke-order sketches, and an Egyptian-vs-modern comparison60
Baseball scoring summary13 sentences and 44 numbers, one per line, each naming who did what, with the running score accumulating at the far right to the final result61

The baseball paragraph's typographic trick, worth stealing: one team is set in sans-serif and the other in serif, so team identity is carried by the typeface and never needs a repeated label. A typeface can encode an entire categorical variable without a legend. (See tufte-redundant-encoding.)

General rule: calibrating a display to the least-informed viewer throws away most of its value. A good display makes every audience smarter from the same ink.


§10. Direct Labels vs. Legends — and What the Swiss Maps Actually Show (pp. 62–63)

A legend is a one-time local code the reader must memorize before the chart means anything. Direct labels put the name at the data, in the data's own color, and abolish the round trip.

The legend indictment (p. 63) is aimed at data-analysis software, not at cartography. On the content-hostile side of Tufte's side-by-side comparison of a machine-learning results chart, five series (bagging, random forest, single tree, lasso, lasso-with-interaction) are named in a disordered legend that consumes more than 40% of the display's space, in black rather than in each data-line's color, with the legend order different from the top-to-bottom order of the lines themselves.

"For 50 years, office and data-analysis software have published trillions of legends — coffins of dead conventions — and trillions of impediments to seeing and learning." — Tufte, Seeing With Fresh Eyes, p. 63

On the content-responsive side the labels sit at the lines, in the lines' colors, with the more significant series enlarged, and all the freed space shows data.

Correction: Swiss topographic maps DO use legends (p. 62)

The Swiss maps are the model for direct labeling, not for legend abolition. Tufte's own text on p. 62 states that maps also use legends, and that Swiss topographic map legends are consistent across all Swiss topographic maps — showing measures of altitude, natural colors for surfaces, and place-names scaled to population size. The p. 62 spread reproduces exactly such a legend: trigonometric points, spot height, index contour, lake level, spot height of lake bottom, plus rock-scree and glacier-moraine rendering shown separately at 1:25 000 and 1:50 000.

The virtue Tufte names for direct labels on those maps is exactitude without clutter — the place name sits at the coordinate it names. Legend abolition is a claim about the p. 63 software chart, not about cartography.

Swiss topo map (p. 62)Software chart, content-hostile (p. 63)
Series/feature namesDirectly at their coordinatesIn a legend box
Legend present?Yes — and consistent across every map in the seriesYes — and unique to this one chart
What the legend carriesSymbol vocabulary reused across thousands of maps: altitude measures, surface colors, scaling conventionsA one-time color code for five lines
Learning valueHigh — learn it once, use it for lifeZero — discarded after this chart
Space costSmall strip at the margin>40% of the display

The population-scaled name table (reproduced by Tufte from swisstopo, SWFE p. 62 — 8 rows):

CategoryPopulationType treatment
Townover 50 000BERN (largest roman caps)
Town10 000 – 50 000LUGANO
Municipality2000 – 10 000Sumvitg
Municipalityless than 2000Cressier
Suburbover 2000Cassarate (italic)
Suburb100 – 2000Champfèr (italic)
Hamlet, group of houses50 – 100Le Plan (italic)
Single house, hut—Tritthütte SAC (smallest italic)

(Swiss map material is reproduced in SWFE with permission of swisstopo, BA200098. Cite it to swisstopo via SWFE p. 62 rather than as Tufte's own table.)

The type is doing two jobs at once: it names the place and encodes its population class through size and roman/italic contrast — a data dimension carried by typography, with the legend confirming the convention rather than replacing the labels.

A working case on the same page (p. 63): the FT's cumulative-deaths-by-days-since-10th-death chart (John Burn-Murdoch) labels every country at its line, in the line's color, and adds content-driven annotations directly in the plot field — lockdown dates, why Italy's elderly population raises its death rate, how South Korea's early large-scale testing and tracing worked, a 33%-daily-increase reference slope.

Placement priority for direct labels:

  1. At the line's terminus (rightmost point for a time series).
  2. Adjacent to the most characteristic feature of the series.
  3. Leader line from the nearest uncrowded whitespace to the datum, only when crowded.
  4. Never a disconnected legend box when direct placement is possible.

The New Haven timetable (EI, pp. 104–107) — verified against the source: the standard railroad table devotes only 21% of its area to the times trains actually run. The top is consumed by grids laboring to make three binary distinctions (New York/New Haven, leaving/arriving, weekday/weekend); the 80 arrival and departure times (~410 characters) are crammed into the leftover space; a small table carries 41 inches (104 cm) of rules; column headings repeat three times with 24 AM/PM markers, forcing a serpentine eye path. The redesign sets the times in Matthew Carter's Bell Centennial and replaces the active colon between hours and minutes with a visually quieter dot.


§11. Banking to ~45° — Who Owns the Aspect Ratio (p. 63)

On the content-responsive side of the same spread, Tufte adjusts the redesigned chart's aspect ratio so that local slopes average roughly 45°.

The rhetorical point is authority, not geometry: why should a publisher's style sheet or a coder's default decide the aspect ratio, forcing every data graphic into the same pre-specified box? Aspect ratio is a content decision, and it currently belongs to whoever wrote the plotting library's defaults.

Cited by Tufte to Jeffrey Heer & Maneesh Agrawala, "Multi-Scale Banking to 45°," IEEE Transactions on Visualization and Computer Graphics, 2006, 701–708. (Procedure and worked ratios: tufte-sparklines §7.)


§12. Letter-Coded Captions and Annotations on Annotations (p. 55)

Letter-coded captions are content-hostile — the purest do/don't pair in the chapter. A surgical illustration captioned "A:" and "B:" below the images forces the same decode round trip a legend forces.

PracticeVerdictWhy
Images labeled A and B, meanings listed underneathContent-hostileThe reader must hold the code in memory and travel between figure and caption
Words placed at the thing, letters deleted, a small glyph (a prohibition mark, a star) carrying the verdictContent-responsiveTufte's name for the fixed version is content-hinted typography: direct labels, just like a map

Annotations on annotations (p. 55). A trade group annotated a photograph of a patient's post-operative IV lines with clinical claims. A clinician then annotated the annotations, in a second color, correcting each claim in place. The display then holds a dispute rather than a conclusion: the reader sees the claim, the rebuttal, and their adjacency — which no prose summary would preserve.

Design rule: when evidence is contested, give the rebuttal a distinct color and put it next to the claim it rebuts, not in a separate document. (Developed in tufte-annotation-as-thinking.)


§13. Graphical Sentences That Carry a Data Table (p. 56)

A CONSORT-style participant-flow diagram is analysed as a single content-responsive graphical sentence reporting 22 numbers: 14,692 assessed for eligibility, 5,331 excluded (itemised into seven reasons), 9,361 randomised, 4,678 versus 4,683 by arm, with dropouts itemised per arm.

A metabolic-pathway chart showing only about 5% of known pathways is offered as the same species: like a map, its content is its format — a grid of high-resolution meaning rather than a picture with a caption.

Test for whether you have one: if you removed the numbers from your diagram, would anything remain that a reader could still act on? If yes, the numbers were decoration. If the diagram collapses, the format and the content are the same thing, which is the goal.


§14. Line Weight and Type Weight as Meaning Hierarchy

Weight is a semantic signal: the heavier mark draws more attention, so the heavier mark must carry more meaning. The data line must visually dominate the connecting verticals, grid, and frame (VDQI, p. 186). The visual elegance of 18th- and 19th-century graphics came largely from copper-plate engraving's hair-thin strokes; the heavier lines of 20th-century drafting pens made charts look clumsy by comparison.

ElementRoleInk weight
Data series / measurementPrimary signalHeaviest
Data-point labels, direct annotationsPrimary contextMedium-heavy
Axis labels (named dimensions)Frame contextMedium
Tick marks, grid linesReference structureLight / hairline
Bounding box, axesContainerHairline or none
Legend box borderStructural noiseEliminate if possible
ElementWrong weightCorrect weight
Data values (numbers)Light / thinRegular or medium
Row / column labelsBoldRegular or medium
Table headerExtra-bold capsSmall caps or medium
Footnote / sourceRegularLight or small
Units ("thousands")Regular inlineSuperscript or light suffix

Table rules (EI, p. 55): Tschichold's hierarchy — rules in a table should be graded, with the heaviest reserved for the most structurally significant division and most horizontal rules simply deleted.


§15. Typeface Choice for Data Contexts

CriterionWhy it mattersTest
Legibility at small sizesLabels must read at 6–8pt in tight spacesReadable in a telephone-book column?
Character distinctiveness0/O, 1/l/I, 6/G, 5/S must not blurSet "1Il0O6G5S" at target size
Low ink in regular weightStructure must not compete with dataIs regular weight lighter than the data marks?
Tabular / lining figuresNumbers in columns must alignEqual horizontal space per digit?
Horizontal readabilityAxis labels run left to rightAny optical trick slowing the horizontal scan?

Bell Centennial (Matthew Carter) — designed for telephone books: high-density numeric columns under poor printing, with ink-trap details holding legibility at small sizes (EI, p. 105). The right model for schedules, dense tables, and tight labels. Carter is also Tufte's source for the letterform-anatomy analysis in Visual Explanations (pp. 98–99), where a diagram is remodeled from 24 to 66 direct labels, and for the survey of ten constructed capital-A forms from 1460–1529 with stroke-to-height ratios between 1:12 and 1:8 (.083–.125), measured against the Trajan inscription cut in 112 AD (VE, pp. 112–113).

Albers on all-caps and sans-serif (EI p. 51; VDQI p. 183). Words are read as pictures of words, so differentiated letterforms read more easily; all-capitals is the hardest setting of all because it flattens word-shape; and Albers judged that a blanket preference for sans-serif demonstrates neither historical nor practical competence. Tufte's friendly-type rules (VDQI, p. 183) follow: upper-and-lower case, not all caps; a typeface designed for text, not a display face; and — a rule worth remembering when tooling fights you — lettering may legitimately be done by hand.

Data and text should share a typeface (VDQI, pp. 180–181). Graphics belong in the same visual register as the prose around them; a chart set in the software's default face announces that it was produced elsewhere and merely pasted in.

Sans/serif as a categorical encoding: see §9's baseball paragraph. Alfred Barr's chart of modern-art movements makes the parallel move with size rather than face, scaling each name in proportion to historical importance exactly as a map scales place-names (BE, pp. 64–68).

ContextPreferenceReason
Body text alongside chartsSerifText rhythm matches explanatory prose
Tight data tables, schedulesSerif or humanist sans with tabular figuresLegibility at small sizes
Axis and tick labelsRegular-weight sans or serifMust sit below the data without competing
Numbers in columnsTabular lining figures onlyAlignment is semantic
All-caps directional labelsAvoidLow word-shape distinctiveness
Chart type vs. surrounding textSame familyData and text belong to one document (VDQI pp. 180–181)

§16. Annotation as Evidence, Not Decoration

Annotation is part of the evidential record. An unannotated image is a picture; an annotated one is an explanatory display.

Galileo's satellites (BE): notebook sketches of Jupiter's moons annotated with times, satellite labels, and distances in Jovian radii — words, numbers, scales, and linking lines. Stripped of annotation the sketch is decorative; annotated, it is evidence of orbital motion. (In SWFE, Galileo appears in a different argument — his 1601 handwritten copy of Ausonio's mirror diagram, pp. 64–66. Same figure, distinct argument.)

Scale suppression as failure (BE): Bloch's Ichthyologie (1795) drew 216 fishes all sized to fit identical boxes, silently erasing scale.

"Every fish has both a different and unknown scale of measurement." — Tufte, Beautiful Evidence

Cleveland's objection and Tufte's answer (BE, pp. 118–121). William Cleveland holds that word-labels on data may interfere with assessment of the overall pattern and prescribes keeping notes, keys, and markers out of the data region. Tufte's rebuttal is the strongest defense of direct labeling in the canon: labels are data (naming the dots is what reveals that large dinosaurs have relatively small brains, and what makes plotting errors detectable at all); abandoning labels forces a code, and codes impede learning; segregating nouns from data-dots splits evidence on a verbal/nonverbal distinction that is uncartographic and substantively irrelevant; reducing data resolution moves the display the wrong way. He then concedes Cleveland has a real point, demonstrates it by deleting all 26 labels — and resolves it by design rather than deletion, pushing labels onto a different visual layer in a grayed field with the data-dots clustered in red.

Annotation typeFunctionPlacement
Scale bar / rulerGrounds the image in measurable realityEdge of the subject area
Direct series labelIdentifies without a legend round-tripAt the series terminus or peak
Explanatory calloutMarks an anomaly or featureLeader line from whitespace to feature
Units suffixDisambiguates numeric valuesInline with the axis label
Source lineEstablishes credibilitySmall type, bottom edge

Annotation density test: remove every label. If the display is still fully self-explanatory, the labels were redundant. If it becomes ambiguous, they were annotation-as-evidence — keep them, and layer them rather than deleting them.


§17. Self-Exemplifying Design

A data display must embody the principles it advocates. Tufte's bad-example scatter plot (VDQI) carries a tiny data set under oversized bold-italic axis labels, fat data points, large mid-plot annotations, all-caps labels, heavy ink texture, and lettering that clashes with the surrounding text. The governing criterion (VDQI) is the clear portrayal of complexity: reveal what is genuinely subtle, do not decorate what is simple.

Checklist:

  • Does type-weight hierarchy match the data-importance hierarchy?
  • Do labels sit at the data, or are they exiled to a legend?
  • If a legend exists, does it teach a vocabulary reused across many displays, or a one-time code?
  • Is the typeface legible at the smallest size it appears, and is it the document's own face?
  • Do all numeric columns use tabular figures?
  • Is all type horizontal? (90-degree rotation slows reading)
  • Are content-responsive linebreaks used wherever lists or parallel structure appear — and were they chosen by the author, not the renderer?
  • Is the aspect ratio a content decision, or a library default?
  • Are small data sets shown as numbers, not encoded into areas or colors?
  • Are paragraphs visually distinct enough to be re-found from memory?
  • Where evidence is contested, does a second color carry the rebuttal in place?

§18. Failure Modes

F0 — LittleDataGraphics. Encoding small data into a pie or bar chart; viewers decode areas and colors back to numbers, and the code teaches nothing. Fix: show the numbers. (SWFE p. 60)

F1 — Content-indifferent linebreaks. Wordspaces and breaks adjusted to square off the right margin, destroying syntax. Fix: break at commas, semicolons, full stops. (SWFE pp. 17, 52)

F2 — The legend tax. Series names exiled to a box eating >40% of the display, with order mismatched to the lines and a color code to memorize. Fix: direct labels at the line termini, in data-line color. (SWFE p. 63)

F3 — Letter-coded captions. "A:" and "B:" under the figure, meanings below — a legend by another name. Fix: words at the thing; a glyph for the verdict. (SWFE p. 55)

F4 — Compositor-chosen breaks. Long labels, headers, and titles wrapped by width alone. Fix: author-specified break points at natural phrases and major brackets. (SWFE p. 17)

F5 — Default aspect ratio. A chart's proportions set by the plotting library. Fix: bank local slopes toward ~45° and treat the ratio as content. (SWFE p. 63)

F6 — Inverted weight hierarchy. Heavy type on structure while data values stay light. Fix: lighten structure, strengthen data labels. (VDQI)

F7 — Scale suppression. Omitting measurement scales, turning evidence into illustration. Fix: embed a scale bar at the point of need. (BE)

F8 — Clotted lettering. Oversized, bold, or all-caps labels consuming space without adding information. Fix: reduce to the minimum label set that keeps the display self-explanatory. (VDQI p. 183; EI p. 51)

F9 — Serpentine reading. Layouts forcing a non-linear eye path — the New Haven table's thrice-repeated headings and 24 AM/PM markers. Fix: linearize reading order. (EI p. 104)

F10 — Rotated labels. Vertical y-axis labels. Fix: stretch the graphic horizontally, transpose, or set labels horizontally. (VDQI)

F11 — Active punctuation in dense data. Colons between hours and minutes, brackets around values — each mark is ink competing with numbers. Fix: a quiet separator (a dot, a thin space). (EI p. 104)

F12 — Unanchored annotation. Annotation floating in whitespace with no positional link to its datum. Fix: leader line, or place it adjacent.

F13 — Uniform paragraph styling. Every paragraph typographically identical, defeating retrieval. Fix: idiosyncratic paragraphs, one distinct treatment per evidence type. (SWFE p. 59)

F14 — Deleting labels to reduce clutter. Treating clutter as an attribute of the information rather than a failure of the design. Fix: layer the labels onto a quieter visual level; do not lower the data resolution. (BE pp. 118–121)

F15 — Foreign typeface. Charts set in the software's default face, visually announcing that they were produced elsewhere. Fix: one type family across data and text. (VDQI pp. 180–181)

F16 — The zero degree of graphics. Design-heavy layouts that fragment or subordinate text, delivering impressions instead of the sequenced depth of crafted prose — Giovannini's pathology, cited by Tufte, noting that books about graphic design are notorious for containing almost no text. The capitulation of text to layout is a named failure, not a style. Fix: let the prose argument set the layout; design serves the sentences. (VE p. 148 n. 25)


§19. Concrete Thresholds

Decision pointGuidanceSource
Minimum label size in printBell Centennial readable in telephone-book columns (~6–7pt at 600dpi)EI, p. 105
Direct label vs. leader lineDirect label when whitespace is adjacent to the datum; leader line only when crowdedEI, pp. 100, 105
Line weight, data vs. structureData line heavier than connecting verticals — no fixed ratio, but data must dominateVDQI, p. 186
Graphic proportionTend horizontal; Playfair favored a ~1.4–1.8 width-to-height ratioVDQI, p. 187
Aspect ratio for slope readingAdjust so local slopes average ~45°SWFE, p. 63 (Heer & Agrawala, 2006)
Legend failure thresholdA legend eating >40% of display areaSWFE, p. 63
When a legend is legitimateWhen it teaches a symbol vocabulary reused across a whole map/chart series, as swisstopo's isSWFE, p. 62
Hours/minutes separatorA quiet dot, not an active colon, in dense schedulesEI, p. 104
Useful-ink floor for a tableAim well above the New Haven table's 21% of area spent on actual dataEI, p. 104
Small-data thresholdAny set small enough to encode in a pie or bar is small enough to print as numbersSWFE, p. 60
Content-responsive linebreak cost+2–3 pages in a 500-page book eliminates ~2–5 billion reader impedimentsSWFE, p. 50
Screen space costZero — content-hostile grids on screens are a choice, not a constraintSWFE, p. 53
Categorical encoding without a labelA serif/sans split can carry a whole categorical variableSWFE, p. 61

References

  • Seeing With Fresh Eyes: Meaning, Space, Data, Truth — Edward R. Tufte (Graphics Press, 2020). Chapter 1: p. 14 (KMP code, 75% whitespace), p. 15 (poetry-line critics), p. 17 (the Great Disjunction — control, 1500 years; the math style sheet; the Ioannidis preview), p. 18 (the dishwasher manual's dangling linebreaks), p. 19 (Bringhurst; Cleveland Clinic thought bubbles). Chapter 2, pp. 49–63: p. 48 (Maugham prelude spread, facing the chapter opener), p. 50 (Darwin), p. 51 (Apollinaire, content-possessed), p. 52 (Ioannidis and Maugham re-set), p. 53 (zero-cost space, alignment as signal), p. 54 (Kepes and the white interval, Hokuba 1808), p. 55 (letter-coded captions, annotations on annotations), p. 56 (CONSORT diagram, 22 numbers; metabolic pathways), p. 57 (hyphens vs. dashes, the suspended sentence), p. 59 (idiosyncratic paragraphs, Galileo mid-sentence mode switch, Lucky Jim stacklist), pp. 60–61 (LittleDataGraphics, NWS KSFO, Egyptian vulture, baseball sans/serif), p. 62 (Swiss topo maps, legends and direct labels; the 8-row settlement table, reproduced from swisstopo BA200098), p. 63 (legend >40%, banking to ~45° with Heer & Agrawala 2006, the FT COVID chart). Also p. 67 (spacing for explanatory depth).
  • Envisioning Information — Edward R. Tufte (Graphics Press, 1990): p. 51 (Albers on word-pictures, all-caps, and sans-serif), p. 55 (Tschichold's table-rule hierarchy), p. 62 (Rhein label placement; more ascenders than descenders), p. 86 (Byrne's spacious leading), pp. 104–107 (New Haven timetable, 21% data area, 104 cm of rules, serpentine reading, dot-vs-colon, Bell Centennial / Matthew Carter), pp. 13 and 113 (vertical Japanese script cooperating with layout), p. 118 (Rameau's swirling words as informational, not ornamental). Verified against the source.
  • Beautiful Evidence — Edward R. Tufte (Graphics Press, 2006): pp. 10–17 (annotation as evidence, Galileo's satellites, Bloch's Ichthyologie and the scale-bar redesign), p. 51 (Gill Sans substitution recovering space), pp. 64–68 (Barr's proportional name-sizing), pp. 118–121 (the Cleveland labels debate, Tufte's five counter-arguments, and the layered redesign).
  • Visual Explanations — Edward R. Tufte (Graphics Press, 1997): p. 97 (interpoints), pp. 98–99 (letterform anatomy, 24 → 66 direct labels, Matthew Carter as source), pp. 112–113 (ten constructed A-forms 1460–1529, stroke ratios 1:12–1:8, the Trajan inscription cut 112 AD).
  • The Visual Display of Quantitative Information, 2nd ed. — Edward R. Tufte (Graphics Press, 2001): pp. 177–191 (aesthetics, line weight, lettering, shape of graphics, horizontal labels), pp. 180–181 (data/text integration; one typeface across both), p. 183 (friendly-type rules; Albers on all-caps and sans-serif; lettering may be done by hand).

All passages paraphrased; quoted lines limited to single attributed sentences with page citations. The Swiss settlement table is third-party licensed material passing through two hands — cite swisstopo via SWFE p. 62.

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.