agentleFS
jpoindexterjpoindexter/tufte-skills/skills/tufte-parallelism/SKILL.md

tufte-parallelism

Skill1 starsChanged 4 months agoMITView on GitHub

Use when designing comparison views, before/after UI, diff displays, overlays, synchronized media panels, identification guides, field guides, or any layout that sets two or more things side by side — or when reviewing a comparison for mismatched scales, broken registration, stacked-in-time sequences, or labels relayed through codes.

The file

We split this skill into the rules it gives an agent

9 rules. About 0 are common, 9 are distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈9

    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:

    Every comparison is either *parallel in space* (elements coexist, eye flicks between them) or *parallel in time* (one element replaces …

    line 49
  2. Distinctive:

    Don't — Gratuitous sequence: put A on screen 1 and B on screen 2 with a transition, forcing the viewer to memorize A.

    line 62
  3. Distinctive:

    When shapes are complex and the differences are fine, lay instances directly over one another (like tracing paper) rather than side by side.

    line 78
  4. Distinctive:

    Do — Overlay for fine shape variation: use distinct line weights/values (black vs gray) so each superimposed instance stays separable.

    line 88
  5. Distinctive:

    When channels are genuinely different media (audio, score, text), don't ask the viewer to fuse them from memory — synchronize them.

    line 92
  6. Distinctive:

    Do — Name the thing where it sits: put the label on/at the element.

    line 135
  7. Distinctive:

    If superimposed: shared registration + a value/weight gap so instances stay separable?

    line 165
  8. Distinctive:

    If synchronized: one clock, active item highlighted in every channel, chrome not eating the data?

    line 166
  9. Distinctive:

    If you redrew a proof-carrying diagram, did you preserve its load-bearing invariants?

    line 167

≈ “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: Parallelism

Overview

Parallelism is the visual analog of rhetorical/syntactic parallelism in prose: matched structure across repeated elements turns a pile of pictures into a comparison. Congruent form (same viewpoint, same scale, same orientation, same track) gives the eye a stable baseline so it can read the variation — the actual signal. The chapter's thesis: every comparison design lives or dies on whether like is genuinely set against like, and the default that wins is adjacent in space, not stacked in time. From Tufte, Visual Explanations, Ch. 5 ("Parallelism: Repetition and Change, Comparison and Surprise"), pp.79–103.

§1. The core idea — comparison is the point, congruence is the tool

Parallelism connects visual elements so the mind reads them as a set. It is the picture-version of parallel prose: Gibbon's Decline and Fall drives an argument with four parallel verbs in one sentence (insinuated, grew, derived, erected) and an italicized by whom set against to whom (pp.78–79). The reader compares because the structure is matched.

"The matching of phrase against phrase, clause against clause, lends an unmistakable eloquence to prose." — Altick, quoted in Tufte, Visual Explanations, p.79

Visual parallels do the same in 2-D/3-D, and they inherit the same risks: ellipsis (omitting an element from one side) and deliberately or carelessly faulty parallelism (§9). The guiding aesthetic:

"good form is clear but not a spectacle." — Tufte, Visual Explanations

The perceiving mind is not passive — it actively hunts links, clusters, and matches (Gombrich, The Sense of Order). Good parallel design feeds that hunt; bad design starves or misleads it.

Anchor examples to keep in mind:

  • Degas, Cheval à l'arrêt (pp.79–80) — bronze horse photo beside an x-ray of its wire armature. The comparison reads instantly because the horse is the same size and stance in both views. Matched pose = legible parallel.
  • Torn/restored newspaper magic trick (p.79) — before/after pair joined by arrow-letters; parallel verbs (torn / restored), parallel images.

Tufte's closing inventory of parallel strategies (p.103) is broader than any one mechanism — pairing, orientation, simultaneity, overlap, superimposition, flowing together along a common track, codes, pointer lines, sequence, adjacency, analogy, similar content — and the chapter's final sentence claims that, handled well, "parallelism becomes the poetry of visual information" (p.103).

§2. The mechanisms that build a parallel (p.82)

Parallelism "grows from a common viewpoint that relates like to like." These are the levers; most strong comparison designs stack several.

MechanismWhat it doesConcrete example
Position / adjacencyPuts elements inside one eyespan so they're compared simultaneouslyDegas photo + x-ray side by side
OrientationAligns the axis/stance so shape differences (not pose differences) carry meaningSame horse stance in both views
Overlap (superimposition)Lays one instance over another (tracing-paper style) to expose fine variationCatich's Trajan letters, §5
SynchronizationMoves separate channels together in timeBeethoven CD-companion, §6
Similar contentRepeats subject so only the changed variable stands outApian's eclipse: same earth/moon, changed shape
Common track / scaleRuns many series along one shared dimension88-key piano scale; Salyut time-axis, §7
Common centeringAligns multiple representations on a shared anchor pointMargaret Morris: photo + notation chart centered on the body

Margaret Morris's Notation of Movement (p.82) binds a photograph to an abstract posture-notation by common centering plus short notes — three representations (photo, diagram, words) reading as one.

Parallelism across modalities — Apian (p.86): eight parallel accounts of one astronomical fact — four sentences and four drawings, plus translations — in which text and image restate each other. The parallel runs between words and pictures, not just picture against picture.

Parallel decay — Wilmarth's Gift of the Bridge (p.84): four channels (hand-lettered text, a date series 1963–1971, image, symbol) deteriorate in step across the sequence. Parallelism can carry change over repetitions, not only likeness.

§3. The central decision — spatial vs. temporal parallelism

Every comparison is either parallel in space (elements coexist, eye flicks between them) or parallel in time (one element replaces another; you compare a remembered image to a present one).

Parallel in spaceParallel in time
LayoutAdjacent panels in one eyespanSequence: one image, then the next
Cognitive costLow — exploits our capacity to canvass, sort, contrast, review at a glanceHigh — must hold a remembered image against the current one
Eye motionQuick local flicksDisorienting back-and-forth / reload
Best forMost comparisons; anything you want studiedReveals, surprise, exact-overlay flips (§4)
RiskEats screen/page spaceMemory load; comparison degrades with delay

Default rule: despite the charm of reveals, comparisons are usually more effective when the information is adjacent in space rather than stacked in time (VE p.81). Tufte re-asserts the rule in Beautiful Evidence (pp.63, 158, 185): stacking evidence in time makes context hard to grasp and relationships hard to evaluate — visual reasoning works best when the evidence sits within one eyespan, so this is a cross-book principle, not a one-chapter preference. Reach for time only when the medium forces it or when a reveal/exact-overlay buys something space can't. (For the eyespan constraint that governs how much can actually sit adjacent before this rule stops paying, see tufte-small-multiples §3.)

  • Do — Adjacent default: show before and after together (above/below or left/right) so the difference is read, not recalled.
  • Don't — Gratuitous sequence: put A on screen 1 and B on screen 2 with a transition, forcing the viewer to memorize A. That's a memory test, not a comparison.

§4. Flip-parallelism — flaps for near-simultaneous, in-position comparison (pp.80–81)

Repton's landscape-gardening books (early 1800s) used hinged flaps: lift the flap and the redrawn "after" replaces the "before" in exactly the same position. This is the one case where temporal beats spatial.

Why the flap works when it's done right:

  • Exact registration — old and new occupy the identical footprint, so rapid flipping gives a near-simultaneous comparison with zero eye travel. This is the precise purpose of parallelism, delivered.
  • Local, contoured flap — Repton's flap was small and shaped to the changed object (a portico added to a cottage), which concentrates attention on what changed. A big rectangular flap over the whole scene would dilute that focus.
  • Layered depth — a façade-only redesign yields a reading sequence: cottage on the flap → decoration beneath → the ghost of the old cottage peeking through the new façade. Parallelism of stacked layers.

UI translation: an onion-skin / hold-to-reveal / opacity-slider diff is a flap. It beats a side-by-side only when registration is pixel-exact and the change is localized.

  • Do — Register exactly, scope tightly: flip/overlay only when both states share an identical frame; shape the reveal to the changed region.
  • Don't — Spectacle over clarity: the "magic" of a reveal is not the goal; if the flip disorients or hides the steady-state, a calm side-by-side wins.

§5. Superimposed parallelism — overlap instead of beside

When shapes are complex and the differences are fine, lay instances directly over one another (like tracing paper) rather than side by side.

  • Catich, Trajan Inscription letters (pp.96–98) — the carved Roman caps vary from occurrence to occurrence; Catich overlays one instance (black outline) on another (gray) to prove they were brush-painted as a guide for cutters, not stamped from uniform stencils. (Breaks in the lines = stone damage, not design.) Superimposition makes the variation visible at a glance — "like Repton's before/after but without a flap" (p.96).
  • Newton's three prism diagrams (p.83) — the same finding (a second refraction doesn't change a ray's color; green stays green) shown three times in parallel: Newton's own sketch, then two editors' redraws. The parallel doubles as a cautionary tale — later editors broke Newton's careful equality of entrance/emergence refraction, corrupting the geometry (see §9, diagram corruption).
  • Interpoints (p.97) — Catich pulls the word-separating dots out of the inscription and arrays them in a parallel row so their differing shapes (brush, not stencil) can be compared element-to-element.
  • Wrong/right teaching pairs (p.96) — Catich's brush-stroke diagrams for painting the letter H set a wrong method (marked "?") beside the correct one (marked "!"), comparing stroke sequence, direction, and path in strict side-by-side parallelism. Contrast pairs as a teaching device — directly transferable to documentation and error-message design.
  • Images run into the sentence (p.97) — Rogers and Morison set five small construction-grid diagrams inline with the prose, so the sequence of grids reads as the visual parallel of five words: the tightest possible text-image parallel.

Trajan reference dimensions (p.97): the inscription is ~277 cm wide × ~114 cm tall (109 in × 45 in).

  • Do — Overlay for fine shape variation: use distinct line weights/values (black vs gray) so each superimposed instance stays separable.
  • Don't — Overlay incompatible things: superimposition needs shared registration and a value gap; without both it's a smear, not a comparison.

§6. Synchronized parallelism — separate channels moving together

When channels are genuinely different media (audio, score, text), don't ask the viewer to fuse them from memory — synchronize them.

  • Beethoven Ninth CD-companion (Robert Winter, Voyager, 1989; pp.88–89) — 4th-movement screen: left column = six motifs as played by the orchestra; right column = the same motifs as later sung by soloists/chorus. Click a box, hear that segment. You hear the joy theme in the 4-part strings, then the same theme in the 4-part voices — the on-screen visual parallel and the sound move together, so the teacher's voice-over is unnecessary. The product holds ~1000 images; German text and English translation run in parallel with the music, and reversed (highlighted) text marks the line currently being sung. A fugue is dissected by sounding each line alone, then in full.
  • Frame tax (failure mode): the ornate screen frame consumed ~30% of the low-res display, leaving room for only ~642 characters — print runs 3–50× that typographic density. Chrome that eats the comparison space is a real cost.

UI translation: aligned, scrubber-linked panels (transcript ↔ audio ↔ waveform; code ↔ preview ↔ logs) where one timeline drives all channels and the active element is highlighted in every channel at once.

  • Do — One clock, many channels: drive every channel from a shared cursor/scrubber and highlight the active item in all of them.
  • Don't — Let the frame eat the data: decoration/chrome that crowds out the channels being compared defeats the parallel.

§7. Common-track parallelism — many flows on one shared scale

Multiple series read as a comparison when they ride a single common dimension.

ExampleCommon trackParallel flowsScale numbers
Music ranges (Pierce, p.87)the 88 piano keys (frequency)instrument families, voice ranges, scale notes88 keys, aligned by Hz
Pop/rock streams (Chapple & Garofalo, pp.90–91)time, 1955–1974~24 stylistic streams, ~470 artists; parents→offspring linked, contemporaries listed per year~20% of phone-book type density
Salyut 6 cyclogram (pp.92–95)a common time-scale8 redundant counts of mission progress; day/night contour bands96-day flight, 91-min orbit, ~1500 sunrises/sunsets
Trimetric cubes (Coe, p.85)a 2-D grid of twist × tilt28 cube images + 252 numbers (9 per cube) + legends28 cubes, 252 numbers

The cube matrix shows a bonus effect: the negative (white) space of the number-matrix echoes the shape of the cube-matrix, so visual, textual, and numerical layers all parallel each other — and the legends sit in rigorous parallel beneath both.

The Salyut cyclogram is the extreme case of redundant parallel counting: cosmonauts Grechko and Romanenko redesigned the meter-long chart in orbit so a single view revealed the whole flight — weeks counting up and down, fractions (⅓, ½, ⅔, ¾), and ultimately eight parallel methods of marking time (moon phases, holidays, weekly tick marks, fraction of total, dated red triangles, elapsed days, weeks-to-go, weeks-done). Planned schedule was drawn pre-flight in colored pens; actual activity was overwritten in red in orbit — a planned-vs-actual parallel on one grid.

  • Do — Anchor every series to one shared, labeled scale: a common axis is what converts "several charts" into "one comparison."
  • Don't — Vary the track silently: different baselines/scales per flow (see §9) destroy the parallel even when the layout looks matched.

§8. Direct labels beat codes — codes obstruct parallelism

A code forces the eye through a relay: image → number → number → noun (and back). Each hop breaks the merge of a part with its name.

  • The cost (pp.98–99): Catich's coded letter diagram scattered ~48 numbers around the glyphs — a one-time, illustration-specific key that keeps part and name apart. The redesign replaced the code with direct labels and, freed of the relay, expanded from 24 named letter-parts to 66, in tight image↔noun parallel.
  • When codes are justified (pp.100–101): highly complex data (e.g. geological field maps) or many densely scattered elements (the photo of 146 astronomers, where numbered traced heads are the only practical index). Even there, ~⅔ of such codes can be eliminated by thoughtful direct labeling plus close text-image integration (the Scheiner sunspot diagram, p.101, is the cautionary case: a letter "O" needing a second mark "X" to say it isn't actually a sunspot).
  • The tightest possible parallel — Kratzenstein's vowel pipes (p.101): the 1779 talking-automaton pipes are lettered A E I O U, and each letter both labels its pipe and is the content — the vowel that pipe pronounces. A label that is the datum leaves nothing to relay; Tufte sets it directly against Scheiner's O-that-isn't-a-sunspot code confusion on the same page.
  • Reading aid in coded displays: give the eye landmarks. In the 146-astronomers diagram, distinctive traced heads (a big hat, a profile, a goatee) act as reference points to locate neighbors — redundant signifiers that rescue an otherwise undifferentiated field.
Direct labelsCodes / keys
Eye pathpart ↔ name, in placepart → number → legend → name
Parallelismtight, unifiedbroken by the relay
Memory loadnonehold the key in mind
Use whenmost cases; layout tools now make it easygenuinely dense/complex maps, scattered fields
Riskcrowding if overdonethe "code curtain" between image and meaning
  • Do — Name the thing where it sits: put the label on/at the element; let image and noun touch.
  • Don't — Invent a one-off code: an illustration-specific number key is the most common, least necessary way to break parallelism.

§9. Faulty parallelism — named failure modes (pp.100–103)

Parallelism fails when the structure claims a comparison the content doesn't honor.

  • Stacked-when-space-was-free — using temporal sequence (memory load, eye reloads) for a comparison that would fit adjacently. The most common avoidable error. Fix: put them side by side.
  • Mismatched scales / baselines (p.103) — paired panels meant to be compared but drawn on different vertical scales (the HIV mortality pair gave men 0–65 and women 0–30, making the women's slope look comparable though men's death risk was more than double): the layout looks parallel, the numbers aren't comparable. The visual parallel lies. Fix: force one shared, identical scale across all panels, or label the break loudly. Same page, same chart, second sin: the rate's denominator ("per 100,000") exiled to a footnote — a parallelism failure between a number and its meaning.
  • Mis-caption (broken parallel structure, p.102) — the caption's parallel ("X, left … Y, right") points at the wrong figure. Tufte's case: a New York Times photo captioned "Napoleon A. Chagnon, left" — but the man at left, a Yanomami man in a loincloth, is not Chagnon; the loincloth-wearing man was evidently invisible to the caption writer, who never registered him as a person needing identification. Fix: verify each label maps to the element it claims.
  • Embellishment smuggling (p.102) — a before/after where "after" quietly adds content beyond the stated change. Repton's redesigns add strolling well-dressed visitors, grazing deer, and (in one) nine sailing boats absent from the grim "before." The flap sells improvements the project never delivered. Fix: hold everything constant except the variable under comparison.
  • Diagram corruption (p.83) — careless re-drawing breaks the geometric equality that carried the original proof. Newton kept entrance/emergence refractions exactly equal; later editors didn't, violating optical law while preserving the appearance of the same figure. Fix: preserve the load-bearing invariants when you redraw.
  • Frame/chrome tax (pp.88–89) — decoration consumes the space the comparison needs (Beethoven screen: ~30% lost to the frame). Fix: strip chrome; spend the pixels on the channels being compared.
  • Spectacle over clarity — the reveal/animation impresses but obscures the steady-state or the difference. Fix: clarity first; magic only if it also clarifies.

Effective unparallelism (p.103): the HIV chart holds two unparallelisms — one inaccurate (the scale mismatch above), one effective. Drawing the HIV line solid against dot-dash background causes deliberately breaks line-style parallelism to accent the one series whose slope is the story — purposeful contrast inside a matched frame is good design, not a violation. Parallelism is a baseline you break knowingly, for the data — never silently, for convenience.

§10. Application playbook

  • Comparison views (A/B, variants, plans): adjacent panels, identical scale/axis/crop, identical viewport size, same orientation. One shared legend, direct labels. Highlight only the delta.
  • Before/after UI: side-by-side by default; add an exact-registration overlay (opacity slider / hold-to-reveal / onion-skin) as the flip option for pixel-precise, localized changes. Never let "after" gain content the comparison doesn't own (embellishment smuggling).
  • Diff displays: superimposed (inline) for fine, in-place changes with a value/weight gap between old and new; split (side-by-side) for structural changes you want studied. Keep both columns on the same scroll/line track.
  • Identification guides (field guides, dashboards, symbol keys): direct labels on the specimen; reserve coded keys for genuinely dense plates, and then seed landmark signifiers so the eye can navigate.
  • Synchronized media (player, replay, observability): one cursor drives all channels; mark the active item in every channel simultaneously; spend space on channels, not chrome.
  • Any "compared with what?" surface: pick space over time unless the medium forces time or an exact-overlay flip pays for itself.

§11. Pre-ship checklist

  1. Is this genuinely a comparison? If so, are the items adjacent in space (not stacked in time) unless time is forced or a flip earns its keep?
  2. Is like set against like — same scale, axis, orientation, crop, viewport, viewpoint? (Mismatched scales = silent lie.)
  3. Is only the variable under study allowed to differ? (No embellishment smuggling.)
  4. Are parts directly labeled rather than relayed through a one-off code? (Codes only for dense/complex fields, then ⅔-prunable.)
  5. Does every label/caption map to the element it claims? (No mis-caption.)
  6. If superimposed: shared registration + a value/weight gap so instances stay separable?
  7. If synchronized: one clock, active item highlighted in every channel, chrome not eating the data?
  8. If you redrew a proof-carrying diagram, did you preserve its load-bearing invariants?
  9. Good form, not spectacle — does the design clarify the difference, or just perform?

§12. Cross-book grounding (Beautiful Evidence, Seeing with Fresh Eyes, Envisioning Information)

  • The 1762 contredanse grid (BE pp.32–33) — a dance pamphlet maps 8 movements onto a constant repeated frame, each cell fusing figures, two floor-plans, trace-tracks, and music. Read down for the sequence within a movement, across for the sequence of movements: a two-axis parallel. Its honest defect — floor-tracks with no explanatory legend — is itself a parallelism failure: structure promising a comparison the legend never licenses.
  • The both-views method (BE pp.29, 45) — show the unannotated image beside the annotated one, so the viewer can audit the overlay instead of swallowing it; the plain panel is the control condition, and Tufte says the device should be used far more often than it is. The direct antidote to embellishment smuggling (§9).
  • The redesign sequence as parallel argument (BE pp.118–121) — Tufte sets three versions of one scatterplot side by side (original, labels stripped, redesigned), the failures struck through in red. Holding the data constant across treatments is what proves clutter was a design failure, not a property of the information. His three principles: clutter is a failure of design, not an attribute of information; never fix a visual problem by reducing content-resolution; fix the design instead. Playbook rule: argue a redesign with all states in one eyespan, never sequentially.
  • Small multiples of a failing format (BE pp.174–176) — one cancer-survival table exploded into six slides looks acceptable slide by slide; reassembled on a single page, the format's ~2.9× area waste becomes legible. To critique a sequential format, break its sequence and show it spatially.
  • Broken prose parallelism (SWFE pp.76–77) — Tufte diagrams Isaiah Berlin's fox/hedgehog paragraph as a stack: fox-hedgehog order held twice, then reversed once and left reversed, labeled broken parallelism producing confusion, repaired by a stacklist. Same-order discipline binds prose and panels alike.
  • Three simultaneous reading directions (VDQI pp.158–159) — the 1970/1979 tax slopegraph sets 15 countries on two dated verticals joined by connecting lines, and is meant to be read three ways at once: down each column for rank at one date, across each line for a country's change, and over the whole field for the pattern of changes. Country names are spaced in proportion to their data values, so the label column doubles as a scale. Tufte calls a layout that licenses several ordered readings of one display a viewing architecture — and parallel structure is precisely what keeps more than one reading direction legible at the same time. Compare §7: a common track buys one comparison; matched parallel form buys several.
  • Parallel sequencing beyond VE (EI pp.23–24, 117) — Envisioning Information names "parallel sequencing" on the butterfly redesign, reuses the Java timetable's vertical to array parallel sequences of data, and reads Zorn's dance-notation rows down for simultaneity and across for sequence.

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.