agentleFS
jpoindexterjpoindexter/tufte-skills/skills/tufte-redundant-encoding/SKILL.md

tufte-redundant-encoding

Skill1 starsChanged 4 months agoMITView on GitHub

Use when deciding whether a variable should ride on more than one visual channel, or when a display shows symptoms of getting that wrong — hue-only encoding that dies in grayscale or for color-deficient viewers, channels colliding into moire or 1+1=3 noise, a second encoding that buys no robustness or separation, or cyclic data severed at an arbitrary boundary such as midnight.

The file

We split this skill into the rules it gives an agent

5 rules. About 0 are common, 5 are distinctive, and 0 are 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.

  • 0

    Very specific

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

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    The channel-removal test (run it on every candidate): delete one channel.

    line 44
  2. Distinctive:

    Do:: put your one or two saturated colors on the smallest, most important marks.

    line 137
  3. Distinctive:

    Do:: treat "colorblind-safe palette" as the *start*, then add value + shape so the chart reads with no color at all.

    line 191
  4. Distinctive:

    Which structural marks carry strong value contrast?: Only the data. Everything below goes gray:

    line 228
  5. Distinctive:

    Removal: Delete one channel. Readable → redundant (go to 2).

    line 241

≈ “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

Tufte: Redundant and Multi-Channel Encoding

Overview

Mapping one data variable to more than one visual channel at the same time is defensible exactly when the extra channel buys something the single channel cannot: separation of crowded layers, robustness across media, or access for viewers and conditions where the primary channel fails. Tufte's Envisioning Information never uses the phrase "redundant encoding"; it builds the idea from layering and separation — using distinctions in shape, value, size, and especially color to make differences legible — and from the destructive opposite, the 1+1=3 effect, where overlapping marks breed unwanted negative-space noise. The canonical positive case is the Swiss topographic map, where color does four jobs at once over the same terrain; the canonical negative case is the heavy grid that vibrates against its own data.


§1. The EI Principle — Differences That Make a Difference

Tufte's governing statement for this whole topic is not the data-ink erasing rule from his earlier book; it is the sentence that opens the final passage of the layering chapter (EI p. 65):

"INFORMATION consists of differences that make a difference." — Tufte, Envisioning Information

The job of multi-channel encoding is to manufacture those differences cleanly. The proof is a paired contrast EI sets up directly: two maps of equal detail, opposite legibility.

MapChannels usedResult
Simla, India (U.S. Army, 1954)One visual level — equal value, equal texture, equal color, near-equal shapeUndifferentiated jumble; "a failure to communicate"
Tokyo Prefecture (1884)Shape + value (light→dark) + size + especially color; informative negative spaceSeparated, layered, calm and distinctive

The lesson is not "add channels." It is that legibility comes from deliberate distinctions across multiple channels, kept in proportion to the importance of each information layer. Color is the strongest separating channel — in the IBM copier parts diagram (EI p. 54) color alone differentiates annotation from the 300 small parts it labels — but strength is exactly why it must be rationed (§5).

Do: separate layers by stacking complementary distinctions (e.g., the Tokyo map's water = blue field that also fades away from each outlined edge — color and gradient and edge). Don't: push every layer to the same visual weight; equal value + equal texture + equal color is the Simla jumble.


§2. Three Outcomes — Redundant vs. Layered vs. Collision

"Two channels on the page" resolves to three very different situations. Diagnose which one you have before defending it.

OutcomeDefinitionRemove-one-channel testVerdict
RedundantSame datum on 2+ channelsStill fully readableJustified only if it buys robustness / accessibility / separation (§4, §7, §8)
LayeredDifferent data/layers on different channelsLose a whole layer of meaningAlways justified — this is layering, the core EI move
CollisionTwo channels modulate each other and breed a third markComprehension improves when you drop oneFailure — 1+1=3 noise (§3)

The channel-removal test (run it on every candidate): delete one channel.

  • Chart stays fully readable → genuine redundancy. Keep only if §7 names a purpose.
  • A layer of meaning disappears → it was layering, not redundancy. Keep.
  • Chart gets clearer → the pair was colliding. Cut or replace the channel.

The frequent confusion: designers call layering "redundant" and cut it to save ink, destroying a real layer. Run the test instead of eyeballing.


§3. The 1+1=3 Failure (Albers)

Draw two black lines and a third visual element appears for free: the bright path of negative space between them. Tufte adopts Josef Albers's formula — 1+1=3, or more — as a fundamental principle of information design. One width plus one width touching reads as 2; the same two widths separated by one width of negative space reads as 3. A field of marks therefore breeds an exponential field of negative shapes. This is the governing failure mode of careless multi-channel work: channels that overlap rather than reinforce manufacture noise.

The noise law: the strength of the 1+1=3 effect is directly proportional to the value (light/dark) contrast between figure and ground. Loud black-on-white marks vibrate; the same marks in gray go quiet. Two corollaries Tufte demonstrates:

  • Thinning two sides of every block so each street is bordered by one thick + one thin line deflects the effect (thin lines read as gray).
  • Replacing black edges with gray establishes "serene, motionless edges" that absorb added detail without vibrating.

Concrete costs measured in EI:

  • Marshalling-signals chart: ~460 lamp-whiskers erased; they had read as glowing light and trembling motion — pure 1+1=3.
  • Tukey multi-window plot: doubled-up, tremor-inducing lines consumed 18% of the chart before redrawing.
  • The aircraft-maneuver index: a black-barred list that vibrates so badly Tufte calls it "perhaps the worst index ever designed."

Channels that COLLIDE (do not pair to encode one variable):

PairFailure mode
Heavy black rule + adjacent typeStripy texture fights the type (the NJ Transit timetable's original grid)
Box border + word insideBorder activates negative white space between word and box (the Surgeon General's warning box)
Hue + saturation, both near-equalTwo settings of one variable read as two categories, not one datum twice
Pattern fill + hue on same fieldCross-hatch over color produces moiré at print/screen resolution
Dark grid + data traceSignal caught in the grid (the ECG trace lost in thick graph paper)

Channels that REINFORCE (safe to combine):

PairWhy it works
Color field + contour edgeArea-grouping and edge-detection are different perceptual systems; no interference
Shape + colorShape survives at zero saturation (grayscale, photocopy); color adds recognition speed
Position + direct labelExact value with and without an axis lookup — two access paths
Gray grid + dataA delicate gray line locates the data without competing with it

"Dark grid lines are chartjunk." — Tufte, Envisioning Information

Do: mute structural marks (grids, frames, rules) to gray so the data carries the only strong value contrast on the page. Don't: wrap words in boxes, double trace-lines, or set tables as nets of enclosed numbers — every enclosure adds a vibrating negative shape.


§4. The Swiss Map — Redundant Color Done Right

EI's existence proof that one variable can ride several channels and gain legibility is the Swiss topographic map (Matterhorn sheet, Landeskarte der Schweiz, 1:25,000). Over the same terrain, color performs four distinct jobs Tufte names explicitly:

Use of colorFunctionOn the map
Label (color as noun)Names a categoryDistinguishes water from stone, glacier from field
Measure (color as quantity)Encodes magnitudeAltitude via contour; rate of change via darkening
Represent (color as imitation)Mimics realityRiver blues, shadow hachures
Enliven (color as beauty)Decorates within reasonLifts the terrain "beyond what could be done in black and white alone"

Altitude is encoded redundantly and complementarily: hypsometric color band (immediate gestalt — high vs. low at a glance), contour line (precise value and, by line density, steepness), and darkening/shadow hachure (rate of change). Each channel does what the others cannot. The black-ink-only strip at the bottom of the same map is the control: legible, but flat — it shows what the redundancy bought.

The over-specification payoff: any one channel of the stack is legible; any two are robust; all together survive color print, grayscale, reduced scale, and color-deficient viewing. Redundancy here is fault tolerance, not decoration.

Co-registration (design inference — not a stated Imhof or Tufte rule): what EI actually states is that GEBCO's color fields are delineated by contours labeled with depth measurements (p. 94), and that on the Swiss map the contour-line color changes as the background shifts (p. 81). The inferred discipline, flagged as such: keep color-field boundaries aligned with contour lines — a color zone bleeding across a contour would manufacture a phantom step-edge, the 1+1=3 effect returning as a registration artifact.

The color-budget formula (EI p. 81):

  • A trained eye discriminates ~1,000,000 colors under contrived pairwise comparison.
  • ~20,000 are usable in practice — the limit is visual memory, not discrimination.
  • For encoding abstract information, more than 20–30 colors produce "not diminishing but negative returns."
  • First principle of color in information design:

"Above all, do no harm." — Tufte's first rule for color in information, Envisioning Information

Byrne's dual labeling (EI p. 87) — the book's most explicit warranted redundancy. Tufte's redraw of Oliver Byrne's color Euclid reinstates conventional letter labels alongside Byrne's color coding. Intermingling the two methods does not read as fussy: it speeds recognition, and it lets each viewer choose their own way of linking text to diagram — with both methods likely used together. This is the working model for §11's purpose test: the second channel earns its place because it resolves a real linking ambiguity.


§5. Imhof's Composition Discipline — Mute the Field, Color the Extremes

Redundant color only works on a quiet stage. Tufte adopts Eduard Imhof's color-composition rules (quoted in EI from Cartographic Relief Presentation):

  • Large areas → muted, dull, grayish, or neutral. Strong, heavy, rich, solid colors belong only on the small areas of extremes. Pure bright colors spread over large adjacent areas have "loud, unbearable effects"; used sparingly on a dull ground they produce extraordinary effects. ("Noise is not music," in the internal quotation EI carries.)
  • Set color against a dull background, not bright white — this is what keeps the Tokyo map's colors "both calm and distinctive."

EI's redesigns all enact this:

RedesignMoveEffect
NJ Transit timetableHeavy grid → tiny leader dots reading as gray"A distinction but not a barricade"
Marshalling signalsGray silhouette + small spots of intense saturated color on the lampsColor carries the signal; silhouette recedes
Nolli map of RomeMute the river's heavy inkingBridge names come forward; rippling-water symbolism kept
Tukey statistical graphColor spots track the path of local averagesNoise of parallel lines and white bands swept away

Do: put your one or two saturated colors on the smallest, most important marks; let everything structural fall to gray. Don't: flood a large background with a strong hue and then try to layer data on top — the field shouts over the signal.


§6. Cyclic Redundancy — Gluing the Data Surface (EI p. 107)

The structural (non-color) case for genuine redundancy is cyclic data, treated in EI's "Narratives of Space and Time" chapter (ch. 6, pp. 97–119; the gluing devices sit on p. 107) — not the layering chapter, flagged here for honest provenance. Three devices share one page, and each spends repetition to remove an arbitrary cut, never to add data:

  1. Playfair's Ordnance chart → torus-graphic. William Playfair handled outlying war-spending peaks first by temporarily extending the grid upward — Tufte's analogy is ledger lines in musical notation — and then by topologically gluing the data surface from top around to bottom, the graphical counterpart of octave displacement. The glued result is a torus-graphic. Thick horizontal bars flag war periods, and the topology surfaces a substantive finding: a ratchet effect, with postwar expense never shrinking back to its prewar level.
  2. The 24-hour timetable. A daily graphical schedule can likewise be glued end-to-end onto a cylinder, so the cycle is fully connected and no train runs off the right edge at midnight — or the grid can simply be prolonged a few extra hours on flat paper, as in the Tufte/Druckrey Atlanta–Chicago airplane schedule, to expose the complete cycle.
  3. Recycled globes. A world map may repeat the globe so that, somewhere in the picture, every country and ocean appears whole — which also avoids ethnocentrism, since no region is condemned to the cut. EI's illustration, the Bryan–Cox world-ocean map, is printed wrapped roughly 1⅔ times around; the figure suggests sizing the overlap to keep the longest continuous feature unbroken and no more — an inference from the illustration, not a rule Tufte states.

Attribution note: the graphical train schedule EI credits in this chapter is Charles Ybry's (British patent No. 11,868, 1846). The Paris–Lyon chart commonly attributed to Marey is VDQI material — do not source it to EI.

Do: repeat a cyclic axis when a natural traversal crosses the boundary, keeping the repetition to what closes the cut. Don't: wrap non-cyclic data onto a torus or repeat a full second period when a fraction suffices — that is density without benefit.


§7. Small Multiples — Differentiated Repetition

The small-multiples chapter (EI ch. 4) is redundant encoding's structural cousin: repeat the frame unchanged, vary only the data, so attention lands on differences. "Constancy of design puts the emphasis on changes in data, not changes in data frames." The technique that connects it to this skill: mute the repeated element, bring forward the differentiator.

ExampleRepeated (muted)Differentiator (foregrounded)
Hudson & Manhattan train signalsThe repeated train outlineColor of the signal lights — redrawing mutes the outline so color reads
Birthplaces of Chinese poets (10,086 across 4 dynasties)Base map, panel layout, and red — the constant hue of every circle across all four mapsCircle size alone carries the poet counts; red is not a second magnitude channel but the single foregrounded data layer
Neurometric brain mapsGrid of disk framesLimited focused color, "more effective than strong rainbow colors"
Fly-fisher plateImplicit gridPairs each insect with its fly-fishing imitation

Two cautions EI raises in these pages:

  • Keep comparisons within the eyespan. The poets/sea-goddess maps fail their intended comparison because the panels are "stranded over two preceding pages."
  • Dark surrounds breed white stripes. The neurometric maps' dark borders generate "disruptive white stripes" — 1+1=3 inside a small multiple. Signal locations by quieter means (an implicit grid).

Do: hold the frame constant, render only the data delta, and reserve saturated color for the one variable that changes. Don't: restyle the frame per panel or surround panels in heavy borders — both convert a small multiple into a vibrating grid.


§8. Modern Extension — Accessibility (flagged: post-Tufte)

Tufte's robustness argument (a display that survives grayscale, reduction, and dull backgrounds) generalizes directly to accessibility. The following is current practice consistent with EI, not from the book.

About 8% of males and 0.5% of females have a color-vision deficiency; the common forms collapse red and green. Hue-only encoding is unreadable for ~1 in 12 male viewers — the same failure as Tufte's grayscale photocopy, just in a different channel. The structural fix is identical: ride the variable on a second channel.

The accessible encoding stack (apply in order):

  1. Value (lightness) — must vary monotonically with the data for sequential scales; this is the channel that survives grayscale and CVD. Design the grayscale version first.
  2. Hue — added for speed for color-capable viewers; use perceptually uniform ramps (Viridis/Cividis sequential, Oklab-derived, ColorBrewer qualitative).
  3. Shape / texture — distinct marker shapes, dash patterns, or hatching so categories survive with zero color.
  4. Direct label — names the category in data space; never needs correct color perception.

Contrast floor: WCAG 2.2 AA requires 4.5:1 for normal text and 3:1 for large text (≥18pt, or 14pt bold) and meaningful graphical objects. Apply 3:1 to data marks, not just labels — a pale dot on white fails regardless of hue.

Do: treat "colorblind-safe palette" as the start, then add value + shape so the chart reads with no color at all. Don't: rely on a swapped palette alone — hue stays fragile under glare, cheap screens, and print.


§9. Named Failure Modes

NameDescriptionSymptom
1+1=3 noiseOverlapping/heavy marks breed bright negative shapesStripy grid texture; vibrating index; doubled trace-lines
Gratuitous redundancySecond channel with no robustness/access/separation purposePasses channel-removal test but buys nothing
Channel collisionTwo channels modulate each otherHue+pattern moiré; box-around-word; value+hue phantom category
Undifferentiated jumbleAll layers at one visual levelSimla map — equal value, texture, color, near-equal shape
Over-saturation of the fieldStrong color spread over large background"Loud, unbearable" ground that drowns the signal
Co-registration failureColor boundary misaligned with edgeTopographic color bleeds across contour → phantom step-edge
Cut cycleCyclic data severed at a round boundaryTrain lines break at midnight; current wraps at the antimeridian
Redundancy inflationRepeating more than the longest natural path needsFull second period when a fraction would close the cycle
Color overloadMore than ~20–30 hues for abstract encodingNegative returns; rainbow maps that read slower than focused color
Stranded comparisonSmall-multiple panels split off the eyespanReader cannot compare without paging back and forth

§10. Decision Tables

Single channel, redundant, or layered?

ContextApproachChannelsNotes
Simple scalar, one mediumSinglePosition or valueAdding a second channel is gratuitous redundancy
Crowded layers (map, schematic, table)LayeredShape + value + size + color, by importanceThe Tokyo/Imhof move; co-register edges
Categorical, mixed mediaRedundantShape + hueShape survives grayscale/CVD
Sequential quantitative, print + screenRedundantValue + hueValue must vary monotonically
Cyclical time-seriesRedundant (spatial)Repeat axis to longest pathNo new info; removes a cut
Topographic / density mapRedundantColor field + contour + darkeningSwiss-map model; mute field, color extremes
High-stakes legibility (medical, safety)RedundantPosition + value + shape/labelSurvives single-channel failure
Internal exploratory chartSingleFastest channelRobustness not worth the density

Which structural marks carry strong value contrast? Only the data. Everything below goes gray:

ElementTreatment
Data marks / signalStrongest value contrast; reserve saturated color here
Grid / graph paperGray, delicate; use reverse-printed side if pre-printed dark; throw out if gridded dark on both sides
Frames / panel bordersLight or implicit; avoid dark surrounds (white-stripe noise)
Labels / typeMatch weight to importance; never equal to the data marks

§11. The Warrant Test (run in order)

  1. Removal. Delete one channel. Readable → redundant (go to 2). Layer lost → layering, keep. Clearer → collision, cut the channel.
  2. Interference. Render both channels at the smallest intended size. Any moiré, white stripe, gap, or banding → the pair collides; replace or drop one.
  3. Purpose. Name the service from {separation of crowded layers · robustness across media · accessibility · cyclic continuity}. None applies → erase the redundancy; "looks more complete" is not a purpose.
  4. Field check. Is the second channel's color confined to small extremes on a muted ground, with edges co-registered? If a strong hue floods the background, mute it before shipping.

The governing asymmetry: a warranted channel costs some density and design time; an unwarranted one (or a collision) costs comprehension. In crowded, multi-medium, or high-stakes displays the asymmetry favors the extra channel — placed quietly, on a quiet ground. Everywhere else it favors restraint.


Sources: Edward Tufte, Envisioning Information — ch. 3 "Layering and Separation" (pp. 53–65), ch. 4 "Small Multiples" (pp. 67–79), ch. 5 "Color and Information" (pp. 81–95; the Swiss map and four uses of color pp. 80–81, Byrne's dual labeling p. 87), and ch. 6 "Narratives of Space and Time" (pp. 97–119; the cyclic gluing devices at p. 107). Josef Albers, "One Plus One Equals Three" (1+1=3 effect, quoted in EI). Eduard Imhof, Cartographic Relief Presentation (color-composition rules, quoted in EI). §8 accessibility is post-Tufte current practice (WCAG 2.2; CIE/Oklab perceptual color; CVD epidemiology), consistent with Tufte's robustness argument but not from the book.

Cross-book notes: VDQI ch. 4 (pp. 96–100) carries Tufte's own inventory of legitimate redundancy — context, comparison, aesthetic balance, and cyclical wrap-around — plus the six-way redundant bar; cite it for the general warrant argument. Seeing With Fresh Eyes p. 61 shows a typeface split (sans vs. serif team names inside one baseball data paragraph) carrying an entire categorical variable with no label — a zero-extra-ink second channel. Visual Explanations pp. 92–93 supplies the operational case: the Salyut-6 cyclogram runs eight redundant parallel counts of the same time — a mission-critical display where the redundancy is the point, because a crew reading under load must be able to recover the time from whichever channel they happen to land on. It is the warrant test (§11) passing at its strongest: a real failure mode (misreading time in flight), and each extra channel reducing exactly that ambiguity.

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.