orchestrate-tufte-vdqi

Orchestrate Tufte VDQI

You are the router. Read the request, decide the intent, and invoke the right skill. You are doing this by understanding, not by matching keywords — the previous version was a brittle keyword function and it is gone.

The toolkit (three skills + one VDQI-sourced reference)

• assess-graphical-excellence — evaluate an existing graphic against the nine criteria, name the chartjunk species present (moiré / dreaded grid / duck / decoration), compute the lie factor and compare it to VDQI's catalogue (NYT MPG 14.8, TIME barrel 59.4, etc.), check whether the data deserves a different genre (table for ≤20 numbers, small multiples for many series, range frame instead of bordered scatter), and emit prioritised fixes tagged with remedy (B1–B7), genre to switch to (C1–C10), anti-pattern resemblance, and exemplar to emulate. The default when intent is unclear.
• render-tufte-chart — produce an actual chart file. Ships per-genre scripts: render_line_svg.py (time-series C10), small_multiples.py (C5), quartile_plot.py (C1), range_frame.py (C2, optional C3 dot-dash marginals), plus wrap_html.py for a tufte-css HTML page. The only skill that outputs a chart.
• references/tufte-principles.md (mirrored into both skills) — the canonical Tufte knowledge, source-grounded to VDQI with page citations: Part A nine criteria with numeric anchors, Part B seven remedies, Part C ten chart genres with construction recipes, Part D chartjunk taxonomy, Part E named-failure catalogue (13 dissected real-world graphics with metrics), Part F named-exemplar catalogue (14 praised graphics with copyable moves), Part G compact quantitative defaults.

Routing

• Evaluate / critique ("is this chart any good?", "what's wrong with this?", "is this misleading?") → assess-graphical-excellence.
• Design / build / produce ("make me a Tufte chart of…", "design a clean time-series", "produce the chart") → render-tufte-chart. If the data is currency across multiple years, deflate it first (remedy B7) before rendering.
• Fix / declutter an existing chart ("clean this up", "too cluttered") → chain: assess-graphical-excellence to diagnose and list remedies, then render-tufte-chart to rebuild honoring them. The assessment's remedy tags (B1–B7) are the instructions render follows.
• Unsure → start with assess-graphical-excellence.

Why this shape

Earlier this toolkit had ten skills. Benchmarking showed most encoded a single Tufte idea the model already applies, so separate routing targets only added latency and risk. Assessment and rendering are the two actions that genuinely benefit from a skill. The depth that makes the skills non-generic lives in the principles reference — specifically Parts C–F, the source-grounded genre playbook, chartjunk taxonomy, named-failure library, and named-exemplar library quoted from VDQI by page. Without those, the skills regress to "Claude paraphrasing Tufte from memory"; with them, the model has Tufte's specific vocabulary and worked numbers to reach for.