Publication-Quality MATLAB Figures

matlab-plot-skill is a reusable skill for Codex and Claude Code that fixes messy MATLAB figures: cramped layouts, unreadable titles, awkward legends, bad white space, and weak PDF output for LaTeX or Overleaf.

The key rule is simple: the agent must not stop after writing plotting code. It must generate the figure, export it, read the rendered figure itself, and iterate until the visual result is no longer a mess.

Find This Repo If You Searched For

• how to make MATLAB figures look professional
• publication-quality MATLAB plots
• fix messy subplot layout in MATLAB
• export MATLAB figure to PDF for LaTeX
• MATLAB figures for Overleaf
• improve MATLAB legend, title, spacing, and white space

What Problem This Solves

• multi-panel MATLAB figures that look cramped or amateurish
• legends that steal too much plot area
• titles that are too long or too small after PDF scaling
• inconsistent colors and symbols across panels
• plots that look acceptable in MATLAB but bad once embedded in a paper

What This Skill Enforces

• MATLAB-first figure cleanup for papers, appendices, and slides
• vector PDF export for LaTeX and Overleaf workflows
• better layout choices for multi-panel figures
• explicit render-review-iterate behavior
• reusable MATLAB export helper code
• a quick PNG review path so the agent can read the rendered figure easily

Repository Layout

• matlab-plot-skill/SKILL.md
• matlab-plot-skill/scripts/export_publication_figure.m
• matlab-plot-skill/references/render_review_checklist.md
• matlab-plot-skill/references/matlab_figure_guidelines.md
• matlab-plot-skill/agents/openai.yaml (Codex UI metadata)
• install.ps1
• install.sh
• examples/demo_publication_figure.m
• examples/output/demo_publication_figure.png (committed preview image)
• tools/validate_repo.py
• LICENSE, CITATION.cff

Prerequisites

• MATLAB R2021a or newer. The export helper uses exportgraphics (R2020a+) and an arguments block (R2019b+); the demo and the documented helper calls use name=value argument syntax (WidthInches=8.4), which requires R2021a+.
• For repo validation only: uv (or Python 3 with pyyaml).

Install

Clone the repo first (the installers run from the clone):

git clone https://github.com/hanlulong/matlab-plot-skill.git
cd matlab-plot-skill

Windows PowerShell

Install into both Codex and Claude Code skill directories:

.\install.ps1

Install only into Codex:

.\install.ps1 -Target codex

Overwrite an existing installed copy:

.\install.ps1 -Force

macOS / Linux

./install.sh

Install only into Claude Code:

./install.sh --target claude

Manual Install

Copy the matlab-plot-skill folder into one or both of these locations:

• Codex: ~/.agents/skills/matlab-plot-skill (Codex releases before 2026 used ~/.codex/skills)
• Claude Code: ~/.claude/skills/matlab-plot-skill

Example Invocation

Use prompts like:

Use $matlab-plot-skill to refactor Plots/generate_my_figure.m into a publication-quality PDF figure. Read the exported figure and the compiled paper page, then iterate until the spacing and titles are clean.

Use $matlab-plot-skill to improve this MATLAB appendix figure. Keep the color mapping fixed across panels, distinguish variants with markers, export a vector PDF, and read the generated figure yourself before stopping.

Demo Example

Run this from the repository root (the path is relative to MATLAB's working directory):

matlab -batch "run('examples/demo_publication_figure.m')"

This generates:

• examples/output/demo_publication_figure.pdf
• examples/output/demo_publication_figure.png

The PNG is there for fast review. The PDF is the publication-style export for the paper workflow. Only the PNG (used as the preview above) is committed; the vector PDF is git-ignored and produced locally when you run the demo.

Validation

tools/validate_repo.py is an optional local check (it is not run by CI). Run it before contributing:

uv run --with pyyaml python tools/validate_repo.py .

(Or pip install -r requirements-dev.txt first, then python tools/validate_repo.py ..)

This checks:

• required skill packaging files, including the committed demo PNG
• SKILL.md front matter and that the body still expresses the render-review-iterate workflow
• the export helper still uses exportgraphics vector output, and the demo still calls that helper
• relative markdown links resolve
• openai.yaml skill metadata
• PowerShell install flow (when PowerShell is available)
• bash install flow (when bash is available)

Making Changes

1. If you change the MATLAB code, lint it locally: matlab -batch "checkcode('matlab-plot-skill/scripts/export_publication_figure.m')" (and the demo).
2. If you change the demo, re-run it with MATLAB and re-commit examples/output/demo_publication_figure.png (the committed preview image).
3. Run the validator above before committing.

Notes

• Codex reads agents/openai.yaml for optional UI metadata.
• Claude Code can use the same SKILL.md-based folder structure.
• The installer copies the skill, so re-run it with --force (-Force on PowerShell) after you update the repo to refresh the installed copy.
• Restart your Claude Code or Codex session after installing so the new skill is picked up.