science-plot-formatter

A Claude Code and Codex skill that beautifies matplotlib scripts for scientific paper submission.

Example

Before	After

What it does

Given an existing .py plotting script, a target conference/journal, and the figure's fraction of column width on the page, the skill renders the chart, inspects it critically, rewrites the visual system (figsize, fonts, linewidths, markers, ticks, legend weight), and re-renders to verify the result. It always renders before editing and again after editing — no code is rewritten from equations alone.

Skill

• beautify-chart — the render → inspect → rewrite → re-render → verify loop. The skill looks up the venue's official author guidelines live (column width, body font size, required font family) and derives the target visual system from the figure's final physical width on the page, not from the current code values.

Installation

This repository ships two equivalent skill definitions:

• Claude Code: .claude/skills/beautify-chart/SKILL.md
• Codex: .codex/skills/beautify-chart/SKILL.md

Claude Code

The instructions below install the skill only into the project you run them in. User-level installation (making the skill available in every project on your machine) is intentionally not covered here.

Option A — Clone and copy the skill folder (recommended)

# 1. Get the repo (anywhere outside your project)
git clone https://github.com/preordinary/science-plot-formatter.git

# 2. From your own project's root, copy the skill in
mkdir -p .claude/skills
cp -r /path/to/science-plot-formatter/.claude/skills/beautify-chart .claude/skills/

To update later, git pull inside the clone and re-run the cp -r (it overwrites).

Option B — Download the single SKILL.md file

If you do not want to clone the repo:

mkdir -p .claude/skills/beautify-chart
curl -fsSL -o .claude/skills/beautify-chart/SKILL.md \
  https://raw.githubusercontent.com/preordinary/science-plot-formatter/main/.claude/skills/beautify-chart/SKILL.md

This grabs just the skill file. You lose the git pull update path — re-run the curl to refresh.

Verify the install

After installation your project should contain:

your-project/
└── .claude/
    └── skills/
        └── beautify-chart/
            └── SKILL.md

Start (or restart) Claude Code in that project and ask it to "beautify a matplotlib script for NeurIPS". The beautify-chart skill should be invoked.

Uninstall

rm -rf .claude/skills/beautify-chart

Codex

Codex skills are installed as skill folders under $CODEX_HOME/skills (usually ~/.codex/skills). Copy the Codex-adapted skill folder there:

# 1. Get the repo (anywhere outside your project)
git clone https://github.com/preordinary/science-plot-formatter.git

# 2. Install for Codex
mkdir -p "${CODEX_HOME:-$HOME/.codex}/skills"
cp -r /path/to/science-plot-formatter/.codex/skills/beautify-chart \
  "${CODEX_HOME:-$HOME/.codex}/skills/"

After installation, restart Codex and ask it to use $beautify-chart on a matplotlib script.

To uninstall:

rm -rf "${CODEX_HOME:-$HOME/.codex}/skills/beautify-chart"

Usage

Ask Claude Code or Codex something like:

Beautify path/to/plot.py for NeurIPS 2026. It's one full column wide.

The skill will:

1. Look up NeurIPS 2026 author guidelines (column width, body pt).
2. Render the script headlessly (MPLBACKEND=Agg) and read the PNG.
3. List concrete visual issues at the target page width.
4. Propose a unified plt.rcParams.update({...}) block plus per-call fixes (errorbar capsize/capthick, stale kwargs removed).
5. Apply the edits, re-render, and verify each listed issue is resolved.

Data, colors, colormaps, linestyles, axis limits, and the user's composition are never touched.

Inputs

Input	Example
script_path	/path/to/plot.py
venue	NeurIPS 2026, Nature Communications, ICML 2026
fraction	0.5 (half column), 1.0 (one column), 2.0 (double columns)

You do not need to supply column widths or body font sizes — the skill looks them up from the venue.

Requirements

• Python with matplotlib installed on the machine running Claude Code or Codex.
• The user's script must be runnable end-to-end (data loadable from the script's working directory).