SOURCE ON GITHUB ↗

evidence.md

view on github ↗69 lines · markdown

Evidence base

What the research and practice say, and which rule in the skill each finding backs. Effect sizes are noted where the source reports them; treat practitioner sources (Roam, McCloud, Tufte) as craft knowledge, not experiments.

Why pictures at all

  • Larkin & Simon (1987), "Why a Diagram is (Sometimes) Worth Ten Thousand Words." Diagrams group related information by location, so the reader doesn't search for it, and they make relations perceptually obvious instead of requiring inference. The "sometimes" matters: the gain comes from spatial layout of structure, not from pictures per se. → Draw structure (process, hierarchy, cause), not decoration.
  • Paivio, dual coding theory. Information encoded both verbally and visually is remembered better than either alone. → Pair a short claim with a picture.
  • Mayer, multimedia learning (Cognitive Theory of Multimedia Learning; many experiments). Words + pictures beat words alone (the multimedia principle). His design principles map directly to the skill:
    • Coherence: cut extraneous material. → no decoration.
    • Signaling: highlight the essential. → one accent color, used in caption and visual together.
    • Spatial contiguity: put words next to the parts they describe. → direct labels, no legend.
    • Redundancy: don't show the same long text next to a graphic it narrates. → delete the paragraph the visual replaces.
    • Segmenting: break complex material into learner-paced pieces. → small multiples, step controls, comic panels.
  • Sweller, cognitive load theory. Working memory is small; split attention between a diagram and separate text that must be mentally merged raises load. → integrate labels into the picture.

Which visual encodings people read accurately

  • Cleveland & McGill (1984). Ranked accuracy of judging quantities: position on a common scale > position on unaligned scales > length > angle/slope > area > volume/color saturation. Replicated by Heer & Bostock (2010) with crowdsourcing. → bars and dots by default; avoid pies, bubbles, 3D.
  • Bertin, Semiology of Graphics (1967). Visual variables (position, size, value, texture, color, orientation, shape) and which suit which data types. → choose the variable by data type.
  • Preattentive processing (Treisman; Healey & Enns). Color hue, size, orientation, and motion "pop out" in under ~250 ms if only one thing differs. → one accent makes the point land at a glance; many colors cancel each other out.
  • Gestalt principles (proximity, similarity, enclosure, connection). The eye groups things that are near, alike, boxed, or linked. → layout encodes structure; misplaced proximity implies relations that aren't there.
  • Munzner, Visualization Analysis & Design (2014). What/why/how framework: the task decides the encoding. → step 1 tags each idea's shape before choosing a form.

Memorability, titles, "chartjunk"

  • Borkin et al. (2013), "What Makes a Visualization Memorable?" Visualizations with human-recognizable objects, color, and visual density were more memorable; common charts were less memorable.
  • Borkin, Bylinskii et al. (2015), "Beyond Memorability." Eye tracking on 393 visualizations: titles and supporting text carry the message; the best titles state the takeaway concisely; pictograms do not interfere with understanding and can improve recognition; redundancy (saying it in the title and the graphic) helps. → title = the claim; icons are fine when they label.
  • Bateman et al. (2010), "Useful Junk?" Embellished charts (Nigel Holmes style) were as accurate as plain ones and better remembered after 2–3 weeks. → beauty and memorable form aren't the enemy; meaningless ornament is. This softens Tufte's data-ink rule.
  • Tufte (The Visual Display of Quantitative Information, Envisioning Information, Visual Explanations). Data-ink ratio, small multiples, sparklines, "show the data", graphical integrity (the lie factor). → small multiples over animation; scale honesty.
  • Cairo, The Functional Art / How Charts Lie. Function constrains form; misleading charts via truncated axes, wrong encodings, cherry-picked ranges. → accuracy test.

Narrative visualization and data comics

  • Segel & Heer (2010), "Narrative Visualization: Telling Stories with Data." Seven genres: magazine style, annotated chart, partitioned poster, flow chart, comic strip, slide show, film/video/animation. Most of these are static and cheap; video is one genre of seven. → the forms catalog covers six without video.
  • Hullman & Diakopoulos (2011); Kosara & Mackinlay (2013). Framing, annotation, and sequence shape what readers conclude. → annotated charts; claim titles.
  • Bach, Riche et al. (2017–2018), data comics and "Design Patterns for Data Comics." Panel layouts, transitions, and zooms for presenting data in sequence.
  • Wang, Bach et al. (CHI 2019), "Comparing Effectiveness and Engagement of Data Comics and Infographics." Lab and in-the-wild studies: data comics were preferred for enjoyment and focus and improved understanding and recall versus infographics and illustrated text. A 2024 medical-data study and a 2026 education study (Shan et al., arXiv 2607.02361, 60 students) found comics beat text on recall and insight comprehension. → comic strip form, for explanation not just entertainment.
  • McCloud, Understanding Comics (1993). Closure: readers fill the gutter between panels with the implied action. Amplification through simplification: simple cartoon faces let more readers see themselves in them. → simple consistent characters; 3–6 panels with an arc.
  • Eisner, Comics and Sequential Art. Panel as unit of time; balloons ordered left→right, top→bottom.
  • Neurath & Arntz, Isotype (1920s–40s). Repeated pictograms, one glyph = fixed quantity, for public education. Modern risk-communication research finds icon arrays help people grasp probabilities (e.g., Galesic, Garcia-Retamero & Gigerenzer 2009; Garcia-Retamero & Cokely reviews). → unit chart form; exact counts.

Animation vs static

  • Tversky, Morrison & Bétrancourt (2002), "Animation: Can It Facilitate?" Reviewing the studies, animations rarely beat equivalent static graphics; apparent wins usually came from extra information or interactivity, not motion. Congruence principle: format should match the concept (change → motion is plausible). Apprehension principle: animations are often too fast or complex to perceive accurately.
  • Höffler & Leutner (2007) meta-analysis (26 studies). A medium advantage for dynamic over static visualizations (d ≈ 0.37), largest when the animation shows procedural-motor knowledge and is realistic (e.g., video of a skill). → motion pays for showing motion; it doesn't pay for decorating static ideas.
  • Robertson et al. (2008), "Effectiveness of Animation in Trend Visualization." Gapminder-style animation was enjoyable but less accurate than small multiples and traces for analysis. → small multiples first; animate only when change is the idea, with pause and replay.
  • Heer & Robertson (2007), animated transitions. Staged, simple transitions help people track objects between states. → if animating, stage it and slow it down.

Concept maps vs mind maps

  • Nesbit & Adesope (2006) meta-analysis (55 studies, 5,818 participants) and Schroeder, Nesbit, Anguiano & Adesope (2017) (142 effect sizes, pooled g ≈ 0.58; constructing g ≈ 0.72, studying g ≈ 0.43). Concept and knowledge maps improve retention compared with reading text or lists. Studying a ready-made map has a smaller but real effect.
  • Mind maps (Buzan) have a thinner and more mixed evidence base, and their unlabeled links don't state relations. → the skill prefers concept-map style: labeled links that read as sentences.
  • Novak & Cañas, concept maps. Node–link–node forms a proposition. → verify each triple as a true sentence.

Visual thinking frameworks (practice)

  • Dan Roam, The Back of the Napkin (2008). Six ways of seeing (who/what, how much, where, when, how, why) map to six picture types (portrait, chart, map, timeline, flowchart, multivariable plot). → the shape table's first six rows.
  • Mike Rohde, The Sketchnote Handbook. Simple shapes + type hierarchy + containers + arrows produce clear visual notes fast. → cheap glyphs, few primitives.
  • Bret Victor, "Explorable Explanations"; Nicky Case's explorables; Distill.pub; The Pudding; NYT Upshot. Interactive, reactive diagrams for abstract ideas. Expensive to build well; use for at most one centerpiece per artifact.
  • Systems thinking (Meadows, Thinking in Systems; Sterman). Causal loop diagrams with polarity and loop labels. → causal loop recipe.
  • Lakoff & Johnson, Metaphors We Live By. We understand abstractions through concrete source domains. → metaphor form, with the mapping made explicit and the break point stated.

LLMs drawing visuals

  • LLMs generate diagrams most reliably as code: SVG, Mermaid, Vega-Lite, HTML/CSS. Benchmarks are young: MermaidSeqBench (2025, arXiv 2511.14967) found big gaps between models on Mermaid sequence-diagram correctness; DiagramEval (2025, arXiv 2510.25761) proposes scoring diagrams as graphs (nodes and edges) because image-similarity metrics miss structural errors. → step 3: extract the node/edge/value structure first, draw from it, then re-check the render against it.
  • LIDA (Dibia, Microsoft 2023), Data Formulator (Microsoft Research) generate charts from data + goals by having the model write plotting code, then self-evaluate and repair. → render, look, fix loop (step 6).
  • Common failure modes in model-drawn SVG: overlapping labels, text overflowing boxes, arrows that miss targets, hard-coded colors that vanish in dark mode, coordinates that look plausible but encode nothing. → the verification checklist.
  • Cost: an SVG diagram is a few hundred to a few thousand output tokens. A generated video is orders of magnitude more expensive and can't be fact-checked node by node.
flow-designflow-designwhat this folder doesA progressive design interview: one question per screen, ending in a clickable HTML prototype of a linear user journey (onboarding, checkout, wizard).global
SKILL.md74 lines
loopsloopswhat this folder doesGive an agent a goal, a budget, a cadence and a model. It works one bounded run at a time while you sleep, inside a fence, and hands you four plain sentences in the morning.system
README.md188 lines
init.sh171 lines
install.sh144 lines
loops-spec.md250 lines
MACHINE.md185 lines
settings.json134 lines
prototype-swarmprototype-swarmwhat this folder doesCrawls a multi-screen HTML prototype for undefined click destinations and spawns one generator agent per missing screen, with a handoff packet written by the source screen, until the click graph is covered.global
SKILL.md162 lines
self-maintenanceself-maintenancewhat this folder doesThe one way an unattended run may touch its own loop registry, and the one place a "need from you" becomes a decision you settle with a single command instead of a paste.system
README.md126 lines
loopctl.py599 lines
loops_registry.py122 lines
rootcauserootcausewhat this folder doesUser research synthesis: groups interview notes, support quotes, board screenshots and metrics by root cause, separates symptoms from causes, and drafts the questions to ask next.global
render.py98 lines
SKILL.md50 lines
show-dont-tellshow-dont-tellwhat this folder doesTurns text-heavy docs and abstract ideas into accurate, code-drawn visuals: diagrams, charts, timelines, flows and comic strips that trace back to the source.global
references02 items
evidence.md69 lines
forms.md150 lines
SKILL.md113 lines
pm-strategistpm-strategistwhat this folder doesA product and strategy advisor agent that blends three lenses (build taste, business structure, PM execution) and pressure-tests ideas instead of cheerleading.global
pm-strategist.md51 lines
todotodowhat this folder doesA TODO list for PMs in one markdown file, reachable from Claude Code and from Telegram: capture, see what's due, triage, and hand items to an agent that asks before it acts.system
README.md59 lines
README.md78 lines
EVIDENCE.MDLINE 001 / 69

drag to scroll