evident-charts

Helps coding agents make clear, honest charts and review them with the checks their tools support. Works in Claude Code, Codex, Cursor, and other agents that support skills.

imageA typical draft ranking raw electric-car counts on a log scale it never mentions, next to the version evident-charts produced after critiquing it: battery-electric share of new cars by country

Left: a typical draft, with raw counts on a log scale, the EU total ranked among its own members, and rainbow colors that mean nothing. Right: the same data after evident-charts critiqued and rebuilt it.

Use

Ask for a chart the way you already do. Your agent can load the skill when making or reviewing a data chart.

Chart sales.csv for a LinkedIn post on which regions grew fastest.
Make one slide showing where our budget went last year.
Critique this chart and fix it.        (attach the PNG, the script, or both)

What it does

• Checks the data for totals mixed with their parts, duplicate rows, and placeholder values.
• Chooses a chart that fits the point you want to make.
• Writes the takeaway as the title and cites the data publisher.
• Checks for overlapping labels, misleading axes, and color problems. It checks matplotlib directly and other libraries through an SVG export.
• Reviews the image, using a separate reviewer when allowed or self-review otherwise, for up to 3 rounds.
• Returns ranked fixes when you ask for a critique, and reports anything it could not check.

Why

Charts from coding agents tend to share the same tells:

• A title that names the topic ("Revenue by region") instead of saying what the data shows
• A legend where labels on the data would fit, and rainbow colors that mean nothing
• Bars that start above 0, or 2 y-axes on a chart
• Numbers in annotations that were never checked against the data
• A guessed source line, or the file name used as the source
• Leftover notes like "TODO" or "confirm" on the image

evident-charts checks these problems with scripts and a review of the rendered image. The checks depend on your agent's tools; skipped checks are reported.

Install

Claude Code

/plugin marketplace add rhiever/evident-charts
/plugin install evident-charts@evident-charts

Cursor and other Agent Skills agents

npx skills add rhiever/evident-charts

GitHub CLI

Set --agent to claude-code, codex, cursor, gemini, github-copilot, or antigravity.

gh skill install rhiever/evident-charts evident-charts --agent claude-code --scope user

Codex

codex plugin marketplace add rhiever/evident-charts
codex plugin add evident-charts@evident-charts

Gemini CLI

The --auto-update flag keeps it current.

gemini extensions install https://github.com/rhiever/evident-charts --auto-update

Requirements

• Python 3.10+ runs the check scripts. matplotlib checks also need matplotlib and numpy.
• SVG layout checks need Chrome or Chromium. Set CHROME_PATH if the checker cannot find it. Your chart library may need an extra package to export SVG.
• Image review needs an agent that can view images. A separate reviewer is optional.

Without Chrome, supported chart-spec checks can still run. Without Python, the agent can review a viewable image but cannot run the scripts. It reports skipped checks and whether it used self-review. An installation of Plotly or Kaleido does not guarantee Chrome is installed.

Update

How the rules work

Each rule has a tag: [E] experimental evidence, [P] practitioner consensus, or [T] taste. Your instructions take precedence. The agent reports requested deviations and any failed checks. Your project's style guide can replace the default look. Citations are in references/sources.md.

Works with

• Libraries: matplotlib, seaborn, pandas .plot, ggplot2, Plotly, Vega-Lite and Altair, and D3. Themes are in assets/themes/.
• Sizes: presets for blogs, X, LinkedIn and Instagram, slides, reports, and phones.

Release package

Build the plugin ZIP with Python; the builder needs no extra packages:

python3 scripts/build_release.py

The ZIP is saved as dist/evident-charts-<version>.zip. It contains the manifests, skill resources, icons, and license, with README assets. Private notes and development files are excluded. See the acceptance checklist for testing an unpacked installation.

The root manifest uses the Agent Plugins format, with Claude and Codex compatibility manifests included. For public submission, upload the ZIP through the OpenAI developer dashboard, then complete its checks and review.

Contributing

Issues and pull requests are welcome. AGENTS.md covers the tests and the rule format; changes are listed in the changelog.

License

MIT. Built by Randy Olson.