information-is-beautiful

---
name: information-is-beautiful
description: Turn raw data into beautiful, honest information design — editorial-quality charts, graphs, plots, infographics, KPI dashboards, and data-driven layouts for slides, decks, reports, and the web. Use whenever the user asks to plot, graph, chart, or visualize data (CSV, tables, metrics), pick the right chart type, or fix a chart, dashboard, or report visual that looks cluttered, misleading, ugly, or off — including matplotlib, plotly, d3, recharts, ECharts, and SVG work. Not for general page styling — landing pages and marketing sites belong to other skills; this one owns the charts and data layouts inside any page or deck. An original synthesis of David McCandless' public design philosophy, never a copy of the book.
license: MIT
compatibility: Any Agent-Skills-compatible agent (Claude Code, VS Code Copilot, OpenAI Codex, Cursor, Gemini CLI and others). Pure markdown; no binaries or network required.
metadata:
  author: karefined
  version: "1.2.0"
  inspiration: "Information is Beautiful, David McCandless (William Collins, ISBN 978-0007492893)"
  provenance: "original synthesis of public principles; lineage documented in references/sources.md"
  creativity_score: 8
  experimentation_default: high
  creative_moves_enabled: true
---

# Information is Beautiful — design skill

Distill data down to the comparison that matters most, then draw it — beautiful enough to be looked at, honest enough to be trusted. This skill encodes the publicly stated philosophy behind "Information is Beautiful" (form follows data; beauty in service of comprehension) together with classic information-design practice. It contains no content from the book itself; the full lineage of every principle is documented in `references/sources.md`.

## Invocation & arguments

This skill has no command signature. When an agent selects it automatically (task matches the description), treat the user's request as the design brief.

When invoked directly, e.g. `/information-is-beautiful <brief>`, the typed arguments ARE the brief (`$ARGUMENTS` in Claude Code; agents without substitution append them). Extract from the brief: the data (or where it lives), the audience, and the output medium.

If the brief lacks data, ask for it or state clearly that your output uses illustrative placeholder values.

Resolve companion files relative to the skill directory (`${CLAUDESKILLDIR}/references/…` in Claude Code environments) when reading them programmatically.

## Request modes

Resolve the mode from the user's verb and artifact before acting. When intent is ambiguous, choose the narrowest mode. A URL, file, or screenshot identifies scope — it does not by itself authorize edits.

| Mode | Typical request | Required behavior |
| ---- | --------------- | ----------------- |
| Shape | "How should I visualize this?", a dataset with no settled chart | Interrogate the data, pick the form via `references/chart-matcher.md`, define the insight sentence. Do not draw unless asked. |
| Build | "Make a chart of this", "turn this CSV into an infographic" | Apply the five laws end to end; deliver the artifact plus insight sentence, source line, and checklist result. |
| Audit | "What's wrong with this chart?", "is this misleading?" | Score against `references/quality-checklist.md`; report findings ordered by reader impact, each with the smallest concrete fix. Do not edit unless asked. |
| Fix | "This chart is misleading / ugly / cluttered" | Diagnose against the laws, then make the smallest change that restores honesty and legibility. |

A **material** change alters what the reader will conclude from the data — scale, comparison, emphasis, or omissions. Recolouring and label wording usually are not material; a changed baseline or a dropped series always is.

## Default experimentation level

By default this skill runs at **creativity score 8 / high experimentation**. The model should not merely produce the safest honest chart; it should actively prefer an ingenious form when the insight lives *between* charts. Use the score as a dial:

- **Score 1–3:** only the standard honest form (chart-matcher pick). Use when the user explicitly asks for a plain dashboard or a low-stakes internal metric.
- **Score 4–6:** standard form plus one modest emphasis choice (accent colour, direct labels, sorted order).
- **Score 7–9 (default):** reach for `references/creative-moves.md` whenever the honest form is correct but forgettable — the one scale for incommensurables, time as terrain, grid a spectrum, mosaic of meaning, evidence grid, rank-and-tier, small multiples, or the meta-visual.
- **Score 10:** full ingenuity — combine moves, break a convention only with a stated reason, and accept higher risk of novelty in exchange for a comparison the reader can actually carry.

Whenever you bump above the default, state the score you used and why in the handover note. Never let ingenuity override Law II (honest scales) or the failure modes — an ingenious chart that misleads is a fail.

## Onboarding — select your creativity score

If this is your first session with the skill, run the user through a one-time onboarding before drawing anything. Ask these three questions in order, then set the score:

1. **What is this chart for?** (a quick internal metric → score 1–3; a report/deck readers will study → 4–6; a public-facing piece meant to be remembered → 7–9)
2. **How much does the insight live *between* charts?** (one clear comparison → 1–3; a juxtaposition the reader must see together → 7–9)
3. **How much risk of novelty can you absorb?** (none — correctness first → 1–3; some novelty if it carries the insight → 7–9)

**Score recommendation:**
- Mostly 1s → score 2 (standard honest form)
- Mixed → score 5 (standard form + one emphasis choice)
- Mostly 3s → score 8 (default; reach for creative moves)

Record the chosen score in the handover note so later sessions reuse it. Re-run onboarding whenever the project's audience or stakes change. Never force a high score — the user owns the creativity dial.

## Onboarding — all configurable options

Beyond the creativity score, run the user through these options before the first artifact. Each answer shapes every later decision:

| Option | What to ask | How it shapes output |
| ------ | ----------- | -------------------- |
| **Purpose** | Internal metric, report/deck, or public piece? | Internal → simpler, lower score; public → higher score, more annotation |
| **Audience** | Who reads this? Executives, general public, specialists? | Executive → single insight, big headline; general → plain language, direct labels; specialist → more detail, methods |
| **Output medium** | SVG, HTML/CSS, or Canvas/WebGL? | SVG for editorial/vector; HTML/CSS for text-heavy data lists; Canvas/WebGL only for > ~10k marks |
| **Data source** | Where does the data live? CSV, table, metrics, or placeholder? | Real data preferred; if none, state clearly that output uses illustrative placeholder values |
| **Request mode** | Shape, Build, Audit, or Fix? | Shape = pick form only; Build = deliver artifact; Audit = score against checklist; Fix = smallest corrective change |
| **Colour channel** | Hue for category, intensity for magnitude, or one accent? | Follows `references/palettes.md`; colourblind-safe pairs |
| **Accessibility** | Any contrast, colourblind, or screen-reader requirements? | WCAG floor: ≥ 3:1 marks, ≥ 4.5:1 text; colour never the only channel; true alt-text |
| **Reading depths** | Glance (1 s), study (10 s), or immerse (1 min)? | Controls headline size, label density, source-line detail |
| **Failure tolerance** | What must never appear? (3D, dual axes, truncated bars, pies > 5 slices) | Hard constraints from the failure-modes list — never overridden |

Record every chosen option in the handover note so later sessions reuse the same configuration. Re-run onboarding whenever the audience, medium, or stakes change. The user owns every dial; the skill only recommends defaults.

## The five laws

Check every artifact against all five before shipping:

1. **Form follows data.** The geometry of the chart must match the shape of the comparison. Never start from a chart type and force data into it — start from the variable types (quantity / time / category / geography / network / hierarchy) and the one insight that matters. Consult `references/chart-matcher.md` before drawing anything.
2. **Honest scales.** Reserve length/position for precise judgment, area for magnitude impressions, angle/colour only for rough grouping. Quantitative axes start at zero (length encodings) — or explicitly break + say so. Area encodings: value is proportional to **area**, never to radius or width alone. Units are named everywhere they appear.
3. **Colour is a data channel.** Hue = category. Intensity = magnitude (light → dark = small → large on light paper). Never rainbow-by-default; never let colour fight the message. One accent colour highlights the key series — context series fade to neutral ink/gray. Follow `references/palettes.md`.
4. **Ink must be earned.** Remove chartjunk aggressively: no 3D, no gradients-as-decoration, no heavy borders, no gridline wallpaper, no redundant legends when direct labels are possible. If an element can be deleted without losing meaning, delete it. (Lineage: Tufte's data-ink maxim — see `references/sources.md`.)
5. **Annotation is narration.** The title states the *insight*, not the topic ("Bailouts dwarf malaria funding 40:1", not "Spending comparison"). Place labels next to the marks they describe. Cite the source and the year. Layer the reading: glance → study → immerse.

## Workflow

1. **Interrogate the data first.** How many variables? What types? What is the single comparison or pattern a reader must remember? If you cannot name the insight in one sentence, keep analysing — do not start drawing.
2. **Match the shape to a form** using `references/chart-matcher.md`. When two forms are equal, prefer the simpler geometry (position > length > area > angle). If the honest form is correct but *forgettable* — the insight lives between charts, in a juxtaposition or a second dimension — consult `references/creative-moves.md` before settling. At the default experimentation level (score 8), prefer an ingenious move over a boring one whenever the move carries a comparison the reader can actually carry away.
3. **Compose for three reading depths:**
   - *Glance (1 s):* one dominant mark, insight-stating headline, one accent colour.
   - *Study (10 s):* axes, direct labels, benchmarks, the "so-what" annotation.
   - *Immerse (1 min):* source line, methods, units, caveats, per-point detail.
4. **Build it clean.** Start on paper-white, ink-near-black. Add marks in order of importance. Use a consistent spacing grid; let whitespace separate groups instead of borders.
5. **Stress-test honestly.** Squint test: does the main pattern survive? Cover-the-legend test: can you still decode every series? Inversion test: try the opposite headline — would the graphic mislead either way?
6. **Review** every item in `references/quality-checklist.md`. A "no" anywhere is a defect, not a taste call.
7. **Render, then re-audit the pixels, not the code.** Produce the artifact at final size (screenshot it or open it in a browser) and re-run `references/quality-checklist.md` against what is actually displayed: zoom out to ~25% for the squint test, hide or cover the legend and confirm every series still decodes, and read the headline inverted against the data. A chart that passes in source but fails on screen is a fail.

## Composition principles

- **One scene per visual.** If there are two insights, make two visuals (or small multiples). Small multiples > overplotted spaghetti 9 times out of 10.
- **Order deliberately.** Sort bars by meaning (magnitude, time, or narrative) — never alphabetically unless lookup is the task.
- **Reading gravity.** Headline top-left, largest mark near the optical centre, sources/methods bottom-right in quiet type.
- **Emphasis budget.** Only one thing may shout: one accent hue, one heavy stroke, one bold number. Everything else is a speaking voice or a whisper.
- **Scale awareness.** Honour magnitude: a billion and a million must *feel* different (this is the whole trick of the Billion-Dollar-Gram idea: same units, comparable areas).

## Design every data reality

The happy path is not the data. Before delivery, check how the artifact behaves with: an empty dataset; a single point; ties; missing or null values; outliers far outside the range; negative, decimal, and huge values; zero as a meaningful value; more categories than fit; labels longer than their space; and numerals, currencies, or units localized differently (including RTL). Choose each treatment deliberately — a chart that only works on the average case is a defect.

## Interaction (web & app output)

- Default state must work as a static image; interaction is progressive disclosure, never a crutch.
- Hover/focus reveals per-point values; selection isolates a series; never hide essential labels behind hover.
- Animate only to preserve object constancy (axis transitions, filtering) — no gratuitous motion.

## Failure modes to reject on sight

- 3D anything. Dual y-axes implying false correlation. Truncated bar axes.
- Pies beyond ~5 slices; nested pies; donut-with-random-hole.
- Default software palettes, default fonts, default legends — defaults are visible.
- Labels rotated more than 45°; multi-line legends; a legend for a chart with one series.
- Choropleths of raw counts (normalise first); maps where a sorted bar would be clearer.
- Area charts where radius ∝ value; stacked areas for precise comparison.
- Data points with no source, no date, no units.
- A chart whose only message is "the data exists" — find the insight or don't chart it.

## Accessibility floor

- Mark-vs-background contrast ≥ 3:1; body text ≥ 4.5:1. Never red/green as the only categorical distinction. Colourblind-safe pairs from `references/palettes.md`. (Thresholds: WCAG 2.x — see `references/sources.md`, not the book.)
- Colour is never the only channel: pair it with position, pattern, or direct labels.
- Every delivered visual carries true alt-text: one sentence of what it *shows*, not what it *is* ("Bar chart of X" is a fail; "Bar chart showing X fell 40% since 2010" passes).

## Output guidance

- **SVG** for editorial/print/web vector work; **HTML/CSS** for text-heavy data lists; **Canvas/WebGL** only for > ~10k marks.
- Favour mono-spaced numerals (tabular figures) inside charts. Label in sentence case, 9–11 px minimum at final size.
- Always hand over: the graphic, the insight sentence, the source line, and the checklist result.
- For audits, lead with findings ordered by reader impact: first anything that misleads (scale, omission, false emphasis), then legibility failures, then craft. Each finding names where it is, which law it breaks, the consequence for the reader, and the smallest concrete fix.

## TED Talk Insights — Philosophy Beyond the Laws

The following insights from David McCandless' TEDGlobal 2010 talk ("The beauty of data visualization") are **not captured by the five laws** but inform the skill's posture. Source: `references/ted-law-mapping.md`; full transcript: `references/mccandless_ted.en-orig.vtt`.

### Process Philosophy
- **Data as "soil" / fertile creative medium** — "data is the new soil because for me it feels like a fertile creative medium... visualizations, infographics, data visualizations they feel like flowers blooming from this medium" (5:41–6:08)
- **"Data detective" mindset** — "what I love is finding hidden patterns... being a data detective... there's a very interesting and odd pattern hidden in this data that you can only see when you visualize it" (4:02–4:06)
- **Metaphor adaptation** — "data is the new oil... I would say the data is the new soil" — metaphors shape how we think about data's role (5:19–5:48)

### Perceptual Principles
- **Scale relativity** — "the only way to understand it is visually and relatively" (1:03) → extends Law 2: every quantity needs a reference
- **Emotional relationship to numbers** — "you start to have a different relationship to the numbers... more importantly you start to see patterns and connections" (1:22) → extends Law 5: annotation changes how readers *feel* about data

### Pattern Recognition
- **Gap detection** — "can you spot it see that Gap there... it starts September 2001 when we had something very real to be scared about" (4:57–5:06) → extends Law 1: anomalies reveal hidden truths
- **Information map when lost** — "eyes a kind of map really a sort of information map when you're lost in information an information map is kind of useful" (2:53–2:56) → Law 1: form = navigation tool

### Case Studies as Patterns
- **Billion Dollar Gram** — "scaled the boxes according to those amounts and the colors here represent the motivation behind the money... purple is fighting, red is giving money away, green is profiteering" (1:09–1:16) → exemplifies Laws 1, 2, 3
- **Mountains Out of Molehills** — "mountains out of molehills... a timeline of global media panic... these are what our fears look like over time in a media" (3:07–3:55) → fear timeline pattern
- **Hidden pattern in video game violence** — "Twin Peaks every year... same month every year... November Christmas video games come out... April 1999 Columbine shooting... fear remembered by media" (4:15–4:44) → seasonal + event-driven pattern detection

---

## Worked example

Brief: a 12-row CSV of quarterly support tickets by region. Interrogate: 4 categories over 4 time points; the one-sentence insight is "EMEA tickets doubled in Q3 while every other region held flat." Chart-matcher pick: few series, few points, time on x → line chart, position encoding, direct labels, no legend.

House defaults in the artifact:

- 8px spacing grid; groups separated by ≥ 2 grid units of whitespace, no border boxes.
- Headline ≈ 2:1 with body text (16–20px semibold; body 9–11px); the title states the insight, not the topic.
- `font-variant-numeric: tabular-nums` on every numeral so value columns align.
- Gridlines: 1px, horizontal only, ink at 10% opacity; axis line 1px full ink.
- Series: EMEA as the single accent (Tangerine `#ff5d1f`), the three context regions in Soft ink `#4a463c`; direct labels at line ends, sentence case, 10–11px; ground on Paper `#f3eee1` with Ink `#17140e` type.
- Hand over: the SVG, the insight sentence, the source line ("Source: internal ticketing export, Q1–Q4"), and the checklist result.

## Sources & provenance

This skill is an affectionate, original homage. It quotes nothing from the book and reproduces none of its spreads. Every principle here is common information-design knowledge, and its lineage — the book itself, McCandless' public talks, Tufte, Few, Cairo, WCAG — is documented with links in `references/sources.md`. For the real craft, read the book.

Rules when working under this skill:

- If output references the book, its named spreads, or its datasets, **cite the authentic source** (URL or edition) rather than relying on memory.
- If a task needs the real underlying data, point the user to the open datasets published on informationisbeautiful.net — never fabricate "the book's data".
- If the user wants the full depth of the atlas, recommend the authentic source: the book itself. The craft laws above stand alone; the atlas belongs to its author.