63 rhiever

evident-charts Skill

>- 在为博客、报告、论文、幻灯片、通讯或社交帖子创建、修改或审阅说明性数据图表时使用。涵盖 matplotlib、seaborn、pandas .plot、ggplot2、Plotly、Altair/Vega-Lite、D3 及 PNG/SVG 图表。也适用于"点评这张图"、"这个图为什么难看"、"让这个图更好看"等带图表图片的请求。不适用于无关编程、软件架构图或非数据类插图。

安装方式:把技能目录放入 ~/.claude/skills/(Claude Code)或在 claude.ai 设置中启用;也可复制右侧安装命令一键添加。

查看源码

技能指令原文(SKILL.md)

evident-charts

Charts that make one point a reader gets at a glance. is this skill's folder: ${CLAUDE_SKILL_DIR} where the host fills it in, else the folder holding this SKILL.md (in claude.ai chat, its copy in the sandbox), never the working directory. Run scripts by absolute path with the project's Python.

Critique, review, or a chart image with a question: follow references/critique.md; otherwise the workflow below.

Host and user instructions

  • User instructions and host requirements take precedence over this skill, including its evidence rules. Preserve an explicit override, explain the consequence briefly, and report the rule deviation and any failed check; do not silently replace the requested chart.
  • Use only available, permitted tools. Python 3.12+ runs the checks; matplotlib checks also need matplotlib and numpy. Palette and SVG scripts use the standard library; SVG layout checks need Chrome/Chromium. Optional helpers or exports may need their own libraries.
  • Without Chrome, run SVG spec checks when a supported spec is available; otherwise skip SVG checks. Without Python, skip script checks. Review a rendered image when the host can view it. Report checks run, failed, skipped, and not checkable, with reasons. A partial check is not a full pass.
  • Use a fresh reviewer only when delegation and vision are available and allowed by the user and host. Otherwise self-review; if no image can be viewed, mark visual review skipped. Do not install tools or delegate outside the host's permissions.

Rules

Rule lines read ID [tag] rule. Break: escape. check: name. [E] experimental evidence: break only with a stated reason. [P] practitioner consensus: break for a clear reason. [T] taste: bend freely. A project's style guide overrides the house look (assets/) and [T] rules; retain integrity guidance unless explicitly overridden. check: names a check in scripts/check_chart.py (scripts/check_svg.py runs a subset of them on other stacks; --list-checks names them); palette means scripts/check_palette.py. Citations: references/sources.md, only when asked why.

Workflow

  1. Analyze. Load the data, print summary stats, confirm units and scope, run the checks below.
  2. Brief in one line: Takeaway: | Audience: | Destination: . The takeaway answers every part of the request (HI-1). Ask at most two questions, only if the takeaway is ambiguous. Presets (assets/presets.json): blog (default), mobile (phone, phone-read newsletters), social (X), social_portrait (LinkedIn, Instagram), slide (deck), report (print, memos, docs).
  3. Choose the form from the takeaway and data shape; open references/choosing.md only when the form is not obvious. It may be a table or one hero number over a sparkline or part bar (stats_size="hero", LY-6). Render alternatives (references/candidates.md) only when unsure between two honest forms.
  4. Build in the project's stack; otherwise Python + matplotlib with scripts/evident.py (other stacks: references/libraries.md). Fit title and subtitle with ev.fits_title and ev.fits_subtitle.
  5. Check with available tools. matplotlib: python /scripts/check_chart.py --dest ; other stacks: export an SVG with live text and run python /scripts/check_svg.py --dest --spec
    (its docstring gives each stack's export). Color-blindness and contrast checks need the render measurement. Fix failures together, except explicit user overrides; retain those findings in the report.
  6. Review the rendered PNG, never the code (references/critique.md, Review loop), using an allowed fresh reviewer or self-review. Fix P0/P1 issues within the user's instructions, re-render, recheck, re-review; stop when none remain to fix, at most 3 rounds. Report unresolved findings and skipped work.
  7. Deliver the format asked for (report may add a PDF): absolute paths to available outputs; the brief; analysis choices a reader should know (any transform, exclusions, thresholds, minimum n); source confidence and anything unconfirmed (SRC-1); alt text when the image is viewable (A1); deviations; checks run, failed, skipped, or not checkable; the one-line review log. Never claim an output was rendered or inspected when it was not.

Analysis checks

  • Use the field's standard comparison: same period last year, calendar months or quarters, named members the audience acts on, standard thresholds; pool, smooth, roll, or rebase only when it would mislead, named in the subtitle (H22, H24).
  • Flag a preliminary last point, never pool it away; compare a partial last period like-for-like (year to date, trailing 12 months, or within the only measure covering it) (H17, H15, H12).
  • Parts sum to the published total; remainder = aggregate minus members; check duplicates, sentinel codes (0, 888, 999, blank), future-dated zero rows (unreleased), split events, and frozen values; compute every number from the data (H11, H24).
  • Show n or intervals for small samples, never rank tiny ones; test survey differences at the publisher's MOE level (H23, U1).

Chart defaults

  • Title answers the question as asked ("what drives X": contribution, not rate; H13); subtitle qualifies it and gives what, units, and when; source line names the real publisher and dataset, a note any data you added; nothing addressed to the user on the chart (TI-1, TI-6, SRC-1, SRC-2).
  • Marks encode the title's measure by position on a common scale; lines only over ordered x (ENC-1, TIME-1).
  • Bars from zero; lines use a range that fits the data (zero only for ratio claims); no dual y-axes, 3D, inverted value axes, or rainbow colormaps (H1, H2, H3, H8, H5, C11).
  • One accent on gray when one element is the story; otherwise color meaningful groups; at most 4 hues (C2, C3).
  • Direct labels over legends (LB-1). Up to about 10-12 marks: value labels replace the value axis, never both; show the numbers the comparison needs; context columns only if needed (VL-1, VL-2).
  • Units on every tick or in the axis title; light value gridlines only; remove ink that carries no information; leave white space; text at least 12 px at display size (TI-7, GR-1, DC-1, TY-2).

Files

Read a reference only when its decision comes up.

| File | Read when |
|---|---|
| references/integrity.md | An analysis, axis, scale, uncertainty, or accessibility question |
| references/text.md | Titles, labels, annotation, layout |
| references/color.md | A color decision beyond the defaults |
| scripts/evident.py | Optional matplotlib helper; its docstring is the usage |
| assets/presets.json, assets/palettes.json | Destination sizes and type scale; house colors |