A computer science PhD student at Yale has made public the Python plotting scripts behind their own papers. The figures ran in Nature Machine Intelligence, ICML, NeurIPS and ECCV. The GitHub repository has more than 12,000 stars. More useful still, they distilled the habits that repeat across those scripts into a Skill. Once your AI agent reads it, it can draw your figures using the same habits. Figures look professional because every single one follows the same set of rules.
The rules at a glance
The figure below is a grouped bar chart from a Nature Machine Intelligence paper in the repository. Blue is the author's own method, and the rest are comparison methods. The legend sits in its own column on the right.

The figure comes from the figures4papers repository. Taking these charts apart, the rules come down to a handful of areas:
| Area | Fixed practice | What the reader sees |
|---|---|---|
| Font and size | Sans-serif type (Helvetica or Arial), size 24 for large figures and 15 to 16 for small ones | Shrink the figure to a column width in a paper and the axis labels are still readable |
| Borders | Remove the top and right frame lines, and give the legend no outer box | A clean image, and the eye lands on the data |
| Colors | Dark blue for your method, green shades for the improved version, pink shades for comparison groups, grey for baselines | Readers find "your" bar before they even read the legend |
| Layout | Several metrics side by side in one wide figure, the legend in its own cell, and the y-axis range tightened to the data | Gaps are obvious at a glance, and the legend never covers the data |
| Printing | Black edges on bars, with diagonal hatching when needed. Export at 300 dpi, or 600 dpi for dense figures, and also export a PDF | Groups stay distinct in black-and-white print, and the figure stays sharp when enlarged |
These practices come from the author's own write-up of their scripts in the repository (for example, most scripts use Helvetica, and most exports are at 300 dpi). We read the documentation and the sample figures. We did not run these scripts ourselves.
Why it works: three reasons
- Color carries meaning before it looks nice. The same color refers to the same thing across the whole paper, so a reader on figure five doesn't need to re-check the legend.
- Remove everything you don't need. Extra frame lines and repeated labels go, and every line that stays earns its place. Reviewers spend little time on each figure, and a clean figure is easier to understand.
- Deliver to print standards. Many people only ever look at their figures on screen, then find out after submission that grayscale printing can't tell the groups apart or that the figure goes blurry when enlarged. These rules export for paper from the start.
Follow along: five steps

Give your AI agent the Scientific Figure Making Skill, then follow these five steps, and have it check each result against the completion standard.
Sign in to unlock this key section
New accounts get 10 free credits, plus 1 more each day you sign in (none on days you don't). Spend 1 credit to unlock and it stays open for you.
Common pitfalls
- The AI fills in missing data. If one cell of data is empty, the agent may quietly estimate a value. Say in your prompt that it may use only the numbers you give it, and check every value against your data once the figure is done.
- Tightening the y-axis too far. A y-axis that starts at 0.85 magnifies small gaps, and reviewers may find that misleading. When you tighten the axis, label the exact values on the bars so readers can judge for themselves.
- Your computer lacks the font. Without Helvetica, the figure falls back to a similar font such as Arial, which usually makes little difference. Some scripts render formulas with LaTeX. If LaTeX isn't installed, tell the agent to turn that option off.
- Copying colors without their meaning. Each color must match who is yours and who is a comparison. Using colors at random can make readers misread the figure.
- Using it commercially. The repository is licensed under CC BY-NC 4.0. Writing papers and giving academic talks is fine, but read the license terms before any commercial use.
Project page: figures4papers. To have your agent draw figures directly under these rules, see the Skill page Scientific Figure Making.
Where it usually breaks
- 01The AI filled in missing data
- 02The y-axis was tightened so much that small gaps looked big
- 03The computer lacks the font or LaTeX
- 04Copied the colors without their meaning
- 05Used it commercially, which breaks the non-commercial license


Comments (0)
No comments yet. Start the conversation.