Display quantitative information

Help the viewer reason from evidence. Prioritise truthful encoding and useful comparisons over cosmetic minimalism. Preserve the package's Tufte-informed methods without turning data-ink or any other maxim into an automatic verdict.

Work from the analytical task

1. Identify the comparison or decision: lookup, trend, relationship, distribution, composition, geography, uncertainty, or monitoring.
2. Inspect grain, units, denominator, groups, missingness, time spacing, transformations, sample size, and uncertainty before plotting. Separate observations from summaries.
3. Choose a display for that task using display selection when needed. Establish intentional aggregation and ordering explicitly.
4. Audit scales and encodings before aesthetics. Check actual data bounds, baseline, transformations, temporal spacing, omitted values, and uncertainty against source data.
5. Deliver the requested chart, code, specification, or critique. Verify the output itself rather than assuming a successful plotting command preserved its meaning.

For a critique, lead with a demonstrated integrity or comparison problem. A sound chart may need no change; do not manufacture an improvement to satisfy a template. For scientific work, retain conditions, uncertainty, calibration, and raw observations where they affect interpretation. For dashboards, align time windows and denominators with the actual decision, not decorative KPI symmetry.

Use helpers deliberately

Resolve SKILL_DIR to this installed directory. Bundled scripts do not live in the target project's scripts/ directory. Read a helper's help before invoking it. The spec audit, display suggestions, lie-factor calculation, and text-fingerprint checks are heuristics or scoped calculations, not a substitute for inspecting data and the rendered graphic. A repetition score does not establish poor writing or AI authorship. Preserve clear consistent terminology instead of varying it randomly.

python "$SKILL_DIR/scripts/suggest_display.py" --csv data.csv --goal auto --format markdown
python "$SKILL_DIR/scripts/audit_visual_display.py" --spec chart.json --format markdown
python "$SKILL_DIR/scripts/render_chart_svg.py" --csv observations.csv \
  --x date --y defect_rate --chart line --x-type date --output new-chart.svg

SVG renderer contract

Choose bar, dot, line, or scatter explicitly. The renderer does not guess a chart type or silently average duplicate observations. It reads at most 10,000 rows from a CSV of at most 16 MiB. New output paths are required; no output overwrite.

• Scatter plots retain all observations and compute both axis limits from those observations, not group means.
• Lines use numeric x values, or --x-type date for unambiguous YYYY-MM-DD dates, with proportional spacing. A blank y value breaks the path; omitted rows cannot establish an unrecorded missing interval. Duplicate x values within a line series are rejected until the caller selects an appropriate aggregation/representation.
• Bars require unique category/group pairs and include zero. Zero values have zero bar height. Dots retain observations; both preserve category encounter order.
• Missing/non-finite/ambiguous numeric values are rejected outside the explicitly supported line-gap case. Locale numbers need deliberate normalisation first.

Optional --group, --title, --metadata, --width, and --height configure the first-pass output. Metadata records input/point counts, domains, gaps, and lack of aggregation. SVG title/description and distinct markers help, but do not certify accessibility. Inspect crowded labels, overplotting, group identification, contrast, and the source/uncertainty annotations needed for the deliverable. Use a full plotting library for intervals, complex dates, dense categories, or publication layout. SVG and optional metadata are separate writes, not an atomic two-file transaction.

References and assets

Load only the relevant material: principles, integrity, redesign, specification, accessibility, examples, and review rubric. Rubric scores organise judgement; they are not empirical accuracy measures. The language/fingerprint material is a review prompt, not an instruction to distort a good existing voice.

Keep the existing chart-spec, critique-note, and handoff templates as optional scaffolds. Preserve units, sources, denominators, uncertainty, and exact numerical meaning through revisions. A cleaner misleading chart is still misleading.

Maintenance

Run python -m unittest discover -s "$SKILL_DIR/tests" -v for the renderer regressions. These test numerical geometry and output handling, not human readability, all retained helpers, or a complete accessibility audit. The existing proprietary licence declaration is retained without a dangling LICENSE.txt link; no new redistribution permission or unavailable historical licence text is invented.