Skip to content

mylesmcook/tufte

v0.1.1ISC

Build and review charts and dashboards to a cited, Tufte-led design standard.

GitHubOpen Website

Tufte

Build and review charts and dashboards to a cited, Tufte-led design standard.

The tufte skill loads a rulebook of about 45 rules. Tufte supplies the philosophy; Cleveland & McGill, Few, Munzner, Brewer, WCAG and the empirical literature fill the gaps on screens, interaction and accessibility. Each rule names the default an LLM tends to produce, a severity, the inputs that can detect it, and a source with its verification status.

Install

Claude Code:

/plugin install tufte --marketplace MylesMCook/mcook-plugins

Codex:

codex plugin add tufte@mcook-plugins-local

Use

  • Build: ask for a chart or dashboard as usual. The skill loads when you build or change a visual, applies the rules, and checks the result before handing it off.
  • Review: ask for a review or critique of a screenshot, HTML/React/Vega-Lite code, a Power BI PBIR or Deneb spec, or a live page. You get at most 7 ranked findings with rule IDs, evidence, fixes and sources, plus a Keep list.
  • Claude Code: the tufte-critic agent gives an independent second opinion on a finished page and edits nothing.

Make it load automatically

In headless tests Claude Code did not load the skill for a plain "build a dashboard" request (0 of 3 runs), even with a strong description. With this line in your global instructions (for example ~/.claude/rules/tufte.md or ~/.claude/CLAUDE.md) it loaded 2 of 2:

Before building, changing or reviewing any chart, graph, dashboard, KPI tile or report page, load the tufte skill and follow it.

Otherwise, name it in the request: "use the tufte skill".

Evidence

Tested on one-line dashboard requests ("Make a dashboard on our support costs this year") with synthetic data containing planted traps:

  • Traps handled rose from 69% to 96% (Haiku) and 73% to 94% (Sonnet). Opus was already at 98%.
  • In a blind review of 20 plain-versus-rulebook pairs, the owner preferred the rulebook build 15 to 5 (p = 0.041).
  • Word budgets (v0.3) cut median page text from 454 to 208 words while traps held.
  • As a reviewer, a stock Opus model already finds the planted problems; the rulebook adds consistent form and sources rather than detection.

Evals, prompts and results live in the source repository (tufte-critic, private).

Limits

  • Rulebook v0.4. Tufte page references marked [B] and five [K] items are not yet checked against the books. The v0.4 type-size and no-clipping rules have not been through a blind review.
  • [P] rules are practitioner judgment and the skill says so.
  • Screenshot-only reviews cannot confirm contrast ratios, scale functions or hover-only content.