The six principles of analytical design from Beautiful Evidence — how to build evidence presentations that assist reasoning rather than decorate reports, why those principles are universal (from the first stone map to any future display), and how to distinguish analytical from decorative design.
インストール
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
The six principles of analytical design from Beautiful Evidence — how to build evidence presentations that assist reasoning rather than decorate reports, why those principles are universal (from the first stone map to any future display), and how to distinguish analytical from decorative design.
Analytical design is design in service of reasoning. Tufte's six principles, derived from Minard's 1869 map of Napoleon's Russian campaign, define what separates displays that assist thinking from those that merely perform it. The core claim: design principles are not arbitrary aesthetic choices — they are derived directly from the cognitive tasks required to reason about evidence. A presentation that violates these principles is not just ugly; it impairs thinking.
Naming convention: Failure-mode labels in this skill (Floating Fact, Forced Flatland, Mode Segregation, etc.) are descriptive teaching names, not Tufte's own terminology. Only quotes and page-cited phrasings reflect his exact words. Use the labels for diagnosis; don't attribute them to Tufte.
§1. The Grand Principle: Design Follows Thinking
The source of all six principles is one meta-principle:
"The principles of analytical design are derived from the principles of analytical thinking." — Tufte, Beautiful Evidence, p. 137
Cognitive tasks drive design decisions. If the intellectual task is comparison, show comparisons. If the task is causal reasoning, show causal structure. This is not a metaphor — it is the literal derivation rule for every choice in information display.
Consequence for practice: Before any design decision, name the intellectual task. The design choice follows from that, not from brand guidelines, template defaults, or what looks impressive. Tufte prescribes that every design process begin by naming the task the display must serve: what does a reader need to do with this information? That question drives which elements to include, how to structure them, and which technologies to employ.
Why the principles are universal (pp. 137–139)
The principles are not tied to a medium, a tool, or an era. Because they come from how humans reason rather than from any technology, they hold across the entire history of evidence display:
Across time: the same rules govern the earliest known maps — scratched into stone roughly 6,000 years ago — and paper, print, screens, and whatever comes next. Tufte's claim is that a principle derived from analytical thinking cannot expire when the production technology changes.
Across minds: the principles would apply even to a hypothetical non-human or alien intelligence presenting evidence, because any reasoning agent must still compare, explain causality, handle many variables, integrate modes, document, and care about content.
Producer/consumer symmetry: the same principles bind the maker of a display and its reader — both are doing the same reasoning. As Tufte puts it, on questions of analytical quality, "we're all in it together." There is no separate, lower standard for "just presenting."
Practical upshot: never excuse a weak display with "it's only a slide / a dashboard / a draft." The standard is set by the reasoning task, which is identical regardless of tool or audience.
Analytical vs. Decorative Design
Dimension
Analytical Design
Decorative / Presentation Design
First question
"What content tasks must this display help with?"
"How can this presentation look good?"
Evidence modes
Integrates words, numbers, images freely
Segregates modes into slides, tables, figures
Multivariate depth
Shows 6+ variables simultaneously when data demands
Strips to 1–2 variables for "clarity"
Comparison
Makes comparisons explicit and visible
Leaves comparisons implicit or verbal
Causality
Shows mechanism, not just correlation
Shows effects without agents or causes
Documentation
Names authors, sources, scales, assumptions
Omits provenance; sources buried in footnotes
Driving force
Content — the substance being explained
Production technology, templates, brand
Test of success
Does it assist thinking about the evidence?
Does it look polished and professional?
§2. The Six Principles
Two epigraphs that frame the principles (p. 122)
Tufte opens the principles with two quotations that set up the whole chapter:
Durkheim, on the categories of understanding. Durkheim held that human reasoning rests on a small set of fundamental categories — among them time, space, number, and cause. Tufte's point: these are exactly the native categories analytical design must serve. The six principles are, in effect, instructions for displaying time, space, number, and cause well.
Matisse, on difference. > "I do not paint things, I paint only the differences between things." — Matisse. This frames Principle 1: the unit of meaning is the contrast, not the isolated object. A display communicates by showing differences.
These two anchors map onto the principles: Durkheim's categories say what to show (number, time, space, cause); Matisse says how meaning arises (through comparison).
Principle 1: Comparisons (p. 127)
State it as:make comparisons, contrasts, and differences explicit and visible.
The central question of all statistical reasoning is: compared with what? Across every form of statistical work — time series, regression, experiment design, database analysis — the central act is answering that question. Every analytical method is, at bottom, a comparison method. The display must make the comparison visible, not leave it implicit.
Minard at the Niemen — read the ratios carefully:
Survival ratio: ~1 in 42 — about 10,000 of the 422,000 who crossed into Russia returned (10,000 ÷ 422,000 ≈ 1/42).
As actually drawn: Tufte notes (p. 127) the two band widths at the Niemen were drawn in a ratio of roughly 1 in 28, not a strict 1-in-42. The graphic encoding is approximate at that crossing, not perfectly proportional. Don't claim the line widths literally are the 1:42 survival ratio — the survival ratio is the fact; ~1:28 is what the ink shows.
Either way, the width of two lines communicates a six-month, hundreds-of-miles catastrophe that no single number conveys.
Do / Don't pairs:
Do
Don't
Show before/after states in the same visual field
Show only the current state without baseline
Provide a reference class or control condition
Report absolute numbers without context
Let scale encode magnitude of difference
Use color alone to signal "different"
Use small multiples for sequence comparisons
Animate sequentially so no two states are visible at once
Place comparanda adjacent or overlapping
Separate comparanda across slides or pages
Failure mode — Floating Fact: a number presented with no comparison reference ("sales were $4.2M"). The reader can't assess significance, trend, or deviation because the baseline was never provided.
State it as:show causality, mechanism, explanation, and systematic structure.
Causal thinking is not optional in serious analytical work. Scientific inquiry is structured by causal laws; medical reasoning about prevention, diagnosis, and intervention requires causal models; policy and reform require knowing which causes to govern — policy decisions are causal decisions about which lever to pull. Displays that show only effects, without causes, cannot support the reasoning they exist to enable.
Minard's temperature graph along the bottom of the map is the causal argument: cold explains why men died during the retreat, not merely that they died. The map gives location; the temperature curve gives mechanism.
Do / Don't pairs:
Do
Don't
Show agents (who or what acts) with arrows or flow lines
Use passive-voice structures that suppress the agent
Include the causal variable even if it adds complexity
Strip to the dependent variable only, for "simplicity"
Draw causal diagrams: A → B → C
Show outcomes with no link between them
Distinguish correlation from mechanism explicitly
Imply causation from co-occurrence
Name the mechanism in the title or annotation
Leave causation buried in the verbal text
Failure mode — Effects Without Causes: a display reports outcomes — what happened — with no agent, mechanism, or sequence. Location without cause; result without actor. The reader cannot reason about what to do because there is no causal model to act on. (For the specific corruption-chapter cases — agentless strategic-plan bullet lists and anti-causal statistics — see §7.)
Principle 3: Multivariate Analysis (p. 130)
State it as:show multivariate data — more than one or two variables at once.
Minard's single map encodes 6 variables: army size (band width), geographic position (2 dimensions, x and y), direction of march, temperature, and date. None of the interesting worlds — physical, biological, human — is bivariate. Reducing to 1 or 2 variables to fit a format is not simplicity; it falsifies the subject's actual complexity.
Two-dimensionality is a constraint of the medium, not of the subject. Presentation technology that enforces flatness trains analysts to think flatly. Because real phenomena are inherently multivariate, showing multiple variables at once should be the default — unremarkable, expected, routine — not a special achievement.
Techniques for escaping flatland:
Layering and separation (transparent overlays, cartographic layering)
Small multiples (one panel per condition, all visible in one view)
Encoding extra variables in size, texture, shape, annotation
Integrating text and numbers directly into graphics
Narrative flow-lines that carry several variables simultaneously
Do / Don't pairs:
Do
Don't
Ask how many variables the phenomenon actually has
Default to a 2-axis chart without examining dimensionality
Use small multiples to add a 3rd or 4th variable
Animate to add a variable (animation hides comparisons)
Annotate data points with words that carry more data
Push extra variables into a legend that must be decoded
Show 6 variables in one display when content demands
Slice into 6 slides, each showing 1 variable
Concrete benchmark (from the source): physical and biological science journals (Nature, Science) publish statistical graphics with a median of more than 1,000 numbers; applied medical research (The Lancet) averages about 45 numbers per graphic. High numeric density is the scientific norm; the low end is a presentation convention, not a content requirement.
Failure mode — Forced Flatland: an inherently multivariate phenomenon (6+ variables) shown with only 1–2, so the appearance of simplicity conceals the real complexity. The reader can't assess fit, variability, outliers, or alternative explanations from the exposed slice alone. (Descriptive label — Tufte's own term for the flat-medium problem is "flatland.")
Principle 4: Integration of Evidence (p. 131)
State it as:completely integrate words, numbers, images, and diagrams.
The evidence doesn't care what mode it arrives in. Segregating it by mode — all images in one section, all tables in an appendix, all text in the body — is organizational convenience that actively impairs reasoning. Maps, the best practice in analytical display, have always integrated multiple modes on a single surface. The distinctions between evidence types matter far less than their combined bearing on the question being investigated, so modes must be assembled where the evidence is needed, not where the production system finds them convenient.
Minard integrates a paragraph of words (title, legend, assumptions), a geographic map (spatial position), flow-lines (army size and direction), temperature data (cause), dates (time), and place names (context) — all within one 25 × 21 inch surface.
Galileo's annotated telescope drawings (1610) show the same move: he wrote observation time, satellite identity, and distance (in Jovian radii) directly onto each sketch, turning still drawings into credible quantitative evidence about satellite motion rather than mere pictures of what the telescope showed.
Do / Don't pairs:
Do
Don't
Place labels directly on data points — words are evidence
Use letter codes keyed to a separate legend
Put the explanatory text with the graphic it explains
Collect figures into a "Figures" section at the back
Annotate a chart with the model equation that fits it
Describe the model in text, show the chart elsewhere
Use sparklines to embed data-lines inline with prose
Write "[see Figure 3]" and break the reading
Show the unmapped and mapped image in sequence
Force the reader to cross-reference between pages
Failure mode — Mode Segregation: a report of all words, with tables appended at the back and images in an exhibit section. The reader must hold evidence from three locations in memory at once; the architecture fights comprehension.
Failure mode — Single-mode Research: an investigation that relies on exactly one kind of evidence (only statistics, or only memoirs, or only economic models) when the subject demands several. The question degrades from "how can this be explained?" to "how can one type of information explain this?"
Principle 5: Documentation (p. 133)
State it as:thoroughly describe the evidence — detailed title, authors and sponsors, data sources, complete measurement scales, and the relevant caveats.
Documentation is quality control, not formality. Undocumented displays are inherently suspect; the presence of documentation signals that someone is taking responsibility for the analysis.
Minard's 1869 map documents all of this on the face of the display:
Documentation item
Minard's answer
What is it about?
Losses in men of the French Army in the Russian campaign, 1812–1813
Who did the work?
Drawn up by M. Minard
Who is that?
Inspector General of Bridges and Roads, retired
Where and when?
Paris, 20 November 1869
Data sources?
5 named sources (Thiers, Ségur, Fézensac, Chambray, and Jacob's diary)
Assumptions?
Troops of Prince Jérôme and Marshal Davout assumed always with the army
Scale (flow-lines)?
1 millimeter = 10,000 men
Scale (map)?
Common leagues of France (Map of Fézensac)
Scale (temperature)?
Degrees of the Réaumur thermometer below zero
Publisher?
Named printer and publisher
Benchmark (from the source): in 13 computer-science books on technical visualizations, only 20% of images had complete scales and labels, and 60% had no scales or labeled dimensions at all. Hubble Space Telescope public images are typically published with no indication of distance, size, or location.
Do / Don't pairs:
Do
Don't
Name the author(s) — people, not agencies
Sign it "Marketing Department" or leave it unsigned
Name every data source explicitly
Write "data from internal records"
Show at least one measurement scale on every graphic
Leave the axis scale implied or omitted
State assumptions and their implications
Bury assumptions in a methodology appendix
Disclose sponsorship and conflicts of interest
Present findings as neutral with no funder named
Date the display
Present it with no temporal context
Failure mode — Anonymous Authorship: corporate and government reports with no named individual author. The absence of names signals evasion of responsibility: when things go wrong, no one is findable. (For undefined units, gamed base years, and time-shifted numbers — corrupt measurement — see §7.)
Principle 6: Content Counts Most of All (p. 136)
State it as:analytical presentations ultimately stand or fall on the quality, relevance, and integrity of their content.
This is the supremacy principle; it overrides the other five. No design technique, visual treatment, or technology can salvage a presentation built on weak, irrelevant, or corrupt content. Conversely, content important enough will be heard even through bad design. Minard's work, Tufte notes, exemplifies the spirit behind excellent analytical graphics: deep knowledge of the subject and genuine caring about the substance.
Minard never names Napoleon on a map of Napoleon's own march — a design choice that reinforces the content priority: memorialize the dead, don't celebrate the surviving celebrity.
The most direct route to a better presentation is therefore stronger content, not a better visual container.
Do / Don't pairs:
Do
Don't
Ask "is this worth showing at all?" before designing
Build the visual container before confirming the content earns it
Start from the substantive question to be answered
Start from the slide template or visual system
Cut content not directly relevant to the question
Pad with tangential data to look thorough
Know the subject deeply — content drives design
Delegate content selection to non-experts
Judge the design by whether it aids reasoning
Judge it by polish or technical sophistication
Failure mode — Chartjunk as Substitute: decorative graphics, 3D extrusion, gradient fills, gratuitous color — visual elaboration that signals effort without adding evidence. The decoration is performing in place of the content.
Failure mode — Technology-Driven Presentation: choosing the display architecture from what the software supports ("how can this presentation use [the latest display technology]?") instead of what the evidence requires. The content question always comes first.
§3. The Analytical vs. Decorative Distinction in Practice
Analytical design is the design of evidence presentations — scientific reports, dashboards, policy briefs, technical manuals, financial disclosures. Decorative / presentation design is the design of communications meant to persuade, brand, or market. They are not mutually exclusive, but their first principles differ:
Analytical Design
Decorative / Presentation Design
Audience task
Reasoning about evidence
Receiving a message or impression
Success criterion
Accurate, efficient inference
Persuasion, recall, aesthetic response
First question
What intellectual task must this support?
What impression should this convey?
Treatment of complexity
Show it — complexity is the subject
Reduce it — complexity fatigues the audience
Treatment of uncertainty
Show it — error bars, intervals, alternatives
Suppress it — uncertainty weakens the message
Integration of modes
Required — evidence crosses all modes
Optional — visual consistency dominates
Documentation
Required — credibility depends on it
Optional or counterproductive (disrupts flow)
The core test: would a skeptical expert, given only the display, be able to assess the quality, provenance, and completeness of the evidence? Analytical design passes this test. Presentation design is not built for it.
§4. Minard's Map as Reference Implementation
Charles Joseph Minard's 1869 Carte Figurative of Napoleon's Russian campaign is Tufte's primary exemplar because it satisfies all six principles at once in a single 25 × 21 inch display:
Principle
How Minard satisfies it
1. Comparisons
422,000 entering vs. ~10,000 returning — survival ratio ≈ 1:42 (bands as drawn at the Niemen ≈ 1:28, p. 127); the Berezina crossing collapses ~50,000 to ~28,000
2. Causality
Temperature curve along the bottom, each point dated; cold explains deaths during the retreat
3. Multivariate
6 variables: army size, geography (x, y), direction, temperature, date — all in one image
4. Integration
Words (title, legend, annotations), numbers (troop counts at each location), flow-map diagram, temperature graphic, geographic map — fully fused
An anti-war image — Minard cared about the dead; the word "Napoleon" never appears
Flow-line scale: 1 millimeter = 10,000 men — the one number needed to read the whole map quantitatively. Note the distinction from §2: the survival ratio is ~1:42; the drawn band widths at the Niemen are ~1:28. The encoding is approximate there, not perfectly proportional.
§5. Failure Modes Catalog
Source column distinguishes the analytical-design chapter (pp. 122–139) from the adjacent corruption chapter (pp. 140–145, see §7).
"Accelerate Revenue Recognition!" with no actor or mechanism
Anti-causal Statistics
2 Causality
140–145
Data mining / factor analysis offered as explanation
Corrupt Measurement
5 Documentation
140–145
Gamed base years, time-shifted or unadjusted numbers
§6. Application Checklist
Before shipping any analytical display, verify each principle:
Comparisons: What is this explicitly compared to? Is the reference visible in the display?
Causality: Does it show mechanism, or only correlation/co-occurrence? Are agents named?
Multivariate: How many variables does the phenomenon actually have? Are fewer shown — and why?
Integration: Are words, numbers, images, diagrams assembled where the evidence is needed, or segregated by mode?
Documentation: Are author(s), sources, scales, date, and material assumptions visible on the display itself?
Content: Would a skeptical expert find the content substantively adequate? Is this worth showing at all?
Universality: Have you excused weakness with "it's only a slide/dashboard"? The standard is set by the reasoning task, not the tool (§1).
§7. Corruption in Evidence Presentations (adjacent chapter, pp. 140–145)
These cases come from the chapter that follows the six principles, not from the principles chapter itself. They are included because each is a canonical, real-world inversion of a specific principle — useful for diagnosis, but cite them to pp. 140–145, not to the principle's own page.
Corruption
Inverts
Diagnostic signal
Agentless effects
2 Causality
Bullet-list strategic plans ("Accelerate Revenue Recognition!") and passive-voice prose (e.g. the 9/11 Commission style) assert outcomes with no actor, action, mechanism, or sequence. Identical-looking bullets can hide radically different causal structures, so the display can't help anyone decide what to do.
Anti-causal statistics
2 Causality
Data mining, factor analysis, and multidimensional scaling crunch a data matrix without testing any causal model — a pile of effects with no causes behind them. Fine for exploration; dangerous when presented as explanation.
Corrupt measurement
5 Documentation
Undefined or imprecise units, tendentiously chosen base years, time-shifting of data (e.g. premature revenue recognition), and inflation-unadjusted money. These are documentation failures before they are dishonesty — the scale, timing, or definition is undisclosed or gamed.
Rule of thumb: corruption is what principle-violation looks like once incentive enters. The fix is the same as the principle's fix — name the agent, test the model, disclose the scale — applied against someone who benefits from the fog.
Source book: /tmp/tufte/beautiful-evidence.pdf
Read pages: "9-45,122-145"
(Read in chunks of ≤20 pages if the range is large)
