# Tufte Critique > Critique a data graphic against a nine-criterion rubric derived from Tufte's VDQI — score it, name the chartjunk species present, compute the lie factor, compare against the book's named-failure catalogue, and return prioritised fixes tagged with the Tufte remedy (B1–B7), genre to switch to (C1–C10), and exemplar to emulate. Use when someone asks whether a chart, graph, plot, or dashboard is any good; what is wrong with it; whether it is misleading or distorted; or how to clean up, declutter, or simplify it. Not for producing a chart file (`tufte-chart`). - Skill: `gnurio/tufte-critique` (Agent Skill, multi-file: 3 files) - Install (CLI): `npx skillmds@latest add gnurio/tufte-critique` - Raw SKILL.md: https://api.skillmd.com/api/skills/gnurio/tufte-critique/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: gnurio (https://skillmd.com/u/gnurio) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/gnurio/tufte-critique --- # Tufte Critique Critique an existing graphic: score it, **name what's wrong using Tufte's vocabulary** (the duck, the dreaded grid, moiré vibration, dimensionality violation, etc.), and hand back the exact remedy plus the genre to switch to and an exemplar to emulate. The output is concrete and source-grounded because the model reasons over a single principles file (`references/tufte-principles.md`) that quotes VDQI by page. Generic "improve data-ink ratio" advice is the failure mode this skill is designed to avoid. ## What you need from the user A description (or image, or file) of the graphic, and ideally its purpose and audience. If something essential is missing, infer reasonably and state the assumption rather than stalling. ## How to critique (six-step workflow) Read `references/tufte-principles.md` first. The workflow uses every part of it. 1. **Score the nine criteria** (Part A). 0–10 each, with a chart-specific observation. Use VDQI's numeric anchors: e.g. data-ink ratio "in several cases … increased from .1 or .2 to nearly 1.0" through editing (p.136); data density 0.15 numbers/in² is "overwrought" (p.162). Unsupported scores are the main failure mode. 2. **Compute the lie factor when proportionality looks suspicious** (Part B, B1). Formula: `(visual change %) / (data change %)`. Acceptable 0.95–1.05. Report the number and **compare to VDQI's catalogue** (Part E) — anchor the verdict in a named case: "this is essentially the 1979 TIME barrel (lie factor 59.4)" or "in the league of the LA Times shrinking-doctor (lie factor 2.8)" rather than a free-floating number. 3. **Identify chartjunk species present** (Part D). Walk the four named offenses and name each one the chart exhibits, citing Tufte's named offender when the resemblance is close: - **Moiré vibration** (cross-hatching, dense stippling, gradients) - **Dreaded grid** (grid darker than the data marks) - **Duck** (decoration drives the chart; visual style > data; dimensionality exceeds data dimensionality) - **Decoration** (ornament that carries no information) 4. **Rank candidate genres and challenge the default** (Part C, Part G). List at least three Tufte genres that could fit the data, ranked by fit. Then apply the **default-challenge rule**: > If your top-ranked genre is also what an unprompted Claude would pick > (line, bar, dot plot, scatter, pie), you MUST do one of: > (a) explicitly justify it by citing what the alternatives lose, OR > (b) reach for a second-line VDQI move (supertable, table-graphic, > sparkline, dot-dash plot, quartile plot, small multiples) and explain > why it is the stronger fit here. > > Quiet defaulting to the obvious chart is the failure mode this rule > exists to catch. Common multi-answer data shapes (flag these in the output so `tufte-chart` knows to emit alternatives): - **1 number / single ratio** → prose statement + tiny inline visual - **≤20 numbers** → supertable + Tufte chart (VDQI p.56) - **Many series of one x** → small multiples + overplotted comparison - **Distributions across groups** → quartile plot + strip plot or histogram - **Bivariate scatter** → range frame + dot-dash marginal variant 5. **Compute the weighted overall score**. Weights: integrity 3×, proportionality 2×, data-ink 2×, typography 0.5×, the rest 1×. 6. **Translate scores into ranked fixes**. Each fix gets up to four tags: - **Remedy**: B1–B7 (the technique). - **Genre**: C1–C10 (the form to switch to). Optional. - **Anti-pattern resemblance**: name a Part-E case the graphic looks like. Optional. - **Exemplar to emulate**: name a Part-F graphic the redesign should resemble. Optional. If the graphic plots a multi-year currency series, check B7. Use `scripts/deflate.py` (requires real CPI values; refuses to guess). **A sound graphic deserves a clean bill.** "The principles should not be applied rigidly or in a peevish spirit" (VDQI p.191) — if the chart is honest and clear, score it high and say so; do not invent problems. ## Output format ``` ## Critique: Context: ### Scores ### Chartjunk species present ### Distortion check Lie factor: — Resembles: ### Genres considered (ranked) 1. — 2. — 3. — Chosen: . Default-challenge: . Multi-render trigger: ### Overall: /10 — ### Fixes (highest impact first) 1. [B?, C?, resembles E?, emulate F?] — 2. ... ``` ## What good looks like All nine criteria scored with chart-specific evidence; chartjunk species named where present; distortion quantified and anchored in a VDQI case; the genre question explicitly considered; recommendations ordered by impact, each tied to a remedy AND (when applicable) a genre switch / a famous case to avoid / a famous case to emulate. The reader should leave with a concrete picture of what to build next, not a list of vague improvements. ## Related skills - `tufte-chart` — rebuild the graphic so it satisfies these criteria, using the per-genre scripts named in your fixes. To fix a cluttered or misleading chart: critique here first, then rebuild there honoring the B-tags.