| name | tufte-critique |
| description | 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`). |
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.
-
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.
-
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.
-
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)
-
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
-
Compute the weighted overall score. Weights: integrity 3×,
proportionality 2×, data-ink 2×, typography 0.5×, the rest 1×.
-
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: <graphic>
Context: <purpose / audience, or stated assumption>
### Scores
<one line per criterion: name — score/10 — chart-specific observation>
### Chartjunk species present
<list any of: moiré, dreaded grid, duck, decoration — each with the detected signature, and Tufte's named offender it most resembles>
### Distortion check
Lie factor: <value or "n/a"> — <interpretation>
Resembles: <named VDQI case from Part E, or "no close analogue">
### Genres considered (ranked)
1. <Part-C genre> — <why it fits this data>
2. <Part-C genre> — <why>
3. <Part-C genre> — <why>
Chosen: <genre>. Default-challenge: <if the chosen genre is what unprompted Claude would also pick, justify here by citing what the alternatives lose; otherwise note "second-line VDQI move — stronger fit than the default chart">.
Multi-render trigger: <one of the data shapes from step 4, e.g. "≤20 numbers ⇒ supertable + chart" — or "none, single canonical render">
### Overall: <weighted score>/10 — <one-sentence verdict>
### Fixes (highest impact first)
1. [B?, C?, resembles E?, emulate F?] — <concrete change>
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.
