tufte-report

Installation
SKILL.md

Tufte Report — Data-Driven Infographic Skill

Create standalone HTML reports that combine editorial narrative with interactive data visualization in Edward Tufte's style: high information density, minimal chart junk, typography-first design.

Design Philosophy

Tufte's core principles drive every decision:

  • Data-ink ratio: every pixel of ink should represent data, not decoration
  • Small multiples: repeat a design to show comparison, not animation
  • Sparklines: word-sized graphics that live inside prose
  • Layering: overview first, then detail on demand
  • Integration: text and graphics share the same visual space (sidenotes, not footnotes)

The report should feel like a well-edited magazine feature — you read it top to bottom, narrative carries you through the data, and every chart earns its space by answering a specific question.

Onboarding — Ask Before Building

Before writing ANY code, ask these questions. Do not proceed until all are answered:

  1. What data sources do you have? (CSV, JSON, SQLite, API endpoint, or raw numbers)
  2. What is the primary question this report should answer? (one sentence — this becomes the title and drives all design decisions)
  3. How many sections do you need? (cap at 8 — push back if more are requested; each section should answer one sub-question)
  4. What's the output format? (standalone HTML file, or embedded component)
  5. Time budget? Provide an estimate:
    • 1-2 sections with tables only: ~200 LOC, ~5 min bypass / ~10 min manual
    • 3-4 sections with 2-3 charts: ~500 LOC, ~15 min bypass / ~25 min manual
    • 5-8 sections with charts + health data + sparklines: ~1200 LOC, ~30 min bypass / ~50 min manual

Scope Protection

This skill enforces hard limits to prevent scope creep:

  • Max 8 sections — if the user asks for more, suggest combining related topics
  • Max 2 chart types per section — a section gets one primary chart and optionally one supporting chart or table. More than that means the section should split
  • Max 3 colors per chart — beyond that, use small multiples instead of rainbow legends
  • No 3D charts, no pie charts, no donut charts — these violate Tufte principles
  • No gratuitous animation — scroll-reveal on enter is fine; spinning, bouncing, or pulsing is not
  • Every chart must have a caption — if you can't write a one-sentence caption explaining what the chart shows, the chart shouldn't exist

When the user asks for something outside these limits, respond with: "That would take the report from [current LOC estimate] to [new estimate]. The extra complexity adds [X] but risks [Y]. Shall I proceed, or can we [simpler alternative]?"

Architecture

report.html (standalone, no build step)
├── Google Fonts CDN (EB Garamond)
├── jsDelivr CDN (Monaspace Argon woff2)
├── jsDelivr CDN (Chart.js 4.x UMD)
├── Inline <style> (design system CSS)
├── Inline HTML (semantic structure)
└── Inline <script> (data + Chart.js configs + sparklines + scroll-reveal)

No build tools, no frameworks, no npm. One file, opens in any browser.

Design System

Read references/design-tokens.md for the complete CSS variables, typography scale, and color palette.

Read references/components.md for the HTML+CSS snippet of every reusable component.

Read references/charts.md for Chart.js configuration patterns and inline SVG sparkline code.

Report Structure Template

Every report follows this skeleton:

1. Title + subtitle + data source tags (monospace, subtle)
2. [Optional] Status dashboard (4-column KPI strip)
3. Overview narrative with inline sparklines + TOC sidebar
4. Summary cards (2-4 KPI tiles with sparklines)
5. Sections (each: state-line → chart+narrative aside → table+narrative aside)
6. [Optional] Decision register (threshold table with status colors)
7. Footer (generation date, sources)

Section Pattern

Each section follows this rhythm:

<h2> with ↑ back-to-top link
<p class="state-line"> — one italic sentence, the takeaway
<div class="aside-container"> — chart on left, narrative on right
<div class="aside-container"> — table on left, interpretation on right

The alternation of chart→narrative→table→narrative creates visual breathing room and prevents "wall of data" fatigue.

Rules for Narrative Text

  • State-lines (the italic intro under each heading): one sentence, max 20 words, states the conclusion not the topic. "HRV down 13%, steps down 42%" not "This section covers health metrics"
  • Aside narratives: 3-4 short paragraphs, each starting with a bold keyword. Written like a newspaper sidebar — facts first, interpretation second
  • Flyouts: reserved for actionable insights or methodology notes. The ✦ symbol marks them as "pay attention"
  • No "tells its own story" or similar filler. Every sentence should contain a number or a decision

Dual-Font Strategy

Context Font Why
All body text, headers, captions EB Garamond Classical editorial feel, excellent readability
All numbers in tables Monaspace Argon Tabular figures align in columns, monospace scannability
Big numbers in cards/dashboards Monaspace Argon Visual weight, distinct from prose
Status indicators, trend percentages Monaspace Argon Precision signaling
Data source tags, code references Monaspace Argon Technical register
Ornament separators (:::) Monaspace Argon with ligatures Programming aesthetic, replaces floral Unicode

Color Principles

Use --ink (near-black) for text, --bg (warm white) for background. Chart colors must be semantically meaningful — don't assign colors randomly:

  • Orange (--spark-claude, #c45a28): primary data stream, effort/work metrics
  • Green (--spark-wispr, #2a7a5a): growth, positive health signals, English language
  • Purple (--spark-social, #5a5aaa): social/communication metrics
  • Blue (rgba(42,80,140)): secondary overlay lines on charts
  • Red (#a02a2a): alerts, negative trends, declining metrics
  • Amber (#c89000): warnings, watch-level signals
  • Green (#2a7a3a): healthy baselines, positive trends

Never use more than 3 colors in a single chart. If you need more, use opacity/saturation variations of the same hue.

Session Lessons (What Goes Wrong)

Based on building the reference report, these are the recurring problems:

  1. Chart.js CDN version: Use @4 not a specific patch version — specific versions may not exist
  2. Chart.js defaults: Set individual properties, never replace entire objects (Chart.defaults.scale.grid.color = '#eee' not Chart.defaults.scale.grid = {color: '#eee'})
  3. Legend circles: Use usePointStyle: false with boxWidth: 8, boxHeight: 8, borderRadius: 4 for true circles. usePointStyle: true creates ovals
  4. file:// protocol: Charts won't load CDN scripts via file://. Always test via localhost
  5. Back-to-back charts: Always separate consecutive charts with narrative, a table, or an ornament. Two charts in a row = "wall of data"
  6. Table overflow on mobile: Wrap in .table-wrapper and add .hide-mobile to secondary columns
  7. Dual-axis charts: Use sparingly — they invite false visual equivalence. Always label both axes clearly
  8. Narrative overreach: Don't claim correlations without computing them. "r = 0.10" is more trustworthy than "strong relationship"

Universal Data Adapter

When the user provides data from any source (CSV, JSON, SQLite, API, raw numbers), normalize it into the standard ReportData intermediate format before generating HTML. This decouples data ingestion from report rendering.

Read references/data-adapter.md for the ReportData JSON schema, field reference, and adapter instructions for each source type.

Workflow:

  1. User provides data → identify source type
  2. Transform into ReportData JSON (ask user for meta.question and desired sections)
  3. Confirm the normalized structure with the user
  4. Generate HTML from the ReportData using the block library

Composable Block Library

Reports are assembled from typed blocks, each with a defined data contract. This replaces ad-hoc HTML generation with a systematic approach.

Read references/blocks.md for the complete block catalog: sparkline-row, kpi-card, trend-chart, data-table, correlation-matrix, narrative, heatmap, strip-chart.

Each block defines:

  • Data contract (what JSON shape it expects)
  • HTML template (copy-paste ready)
  • Composition rules (how blocks pair and sequence)

Preview Server

For iterative development, use the built-in live-reload server:

python3 ~/.claude/skills/tufte-report/scripts/serve.py report.html

Serves on localhost:8042, auto-reloads on file change with scroll position preserved. Zero dependencies — Python stdlib only.

Read references/preview-server.md for details. After generating a report, offer to start the preview server for the user.

Weekly Installs
9
GitHub Stars
137
First Seen
Today
Security Audits