nice-figures

A scientific & blog plotting library built for AI agents. It teaches an agent to generate matplotlib figures in a soft-pastel, research-blog visual register — the kind of plot you see in modern ML/alignment write-ups. Bold sans-serif display titles, scatter overlaid with smoothed trends and shaded confidence bands, signature rounded bars, minimal axes, and ↓better badges. White background by default, so the output is conference- and paper-ready (PDFs embed Type 42 fonts, so they pass IEEE/ACM/NeurIPS font checkers).

Why agent-first matters in practice:

• Recipes, not docs. 16 copy-paste-runnable chart archetypes — every code block is executed in CI, so what the agent copies is never stale.
• A self-verification loop. The skill instructs the agent to render the PNG, look at it, and fix what it sees (legend collisions, unlabeled series, clipped bars) before delivering — the failure modes that "code ran fine" never catches.
• Opinionated defaults over choices. One style, tested helpers (top_legend, plain_log_ticks, soft_colorbar, rounded_bars), and explicit common-mistakes lists, because agents do best when the right thing is the only documented thing.

For Claude Code it ships as a plugin with a skill that triggers automatically; any other agent (Codex, Cursor, …) can use it via AGENTS.md and pip install git+https://github.com/Mapika/nice-figures.

The style at a glance

Multi-panel training curves — scatter under a smoothed trend with a shaded band, minimal axes, and ↓better badges:

Grouped bars with the signature softly-rounded tops, in the warm coral/peach palette:

Kaplan/Chinchilla-style scaling laws — per-run loss curves colored by model size, tangent to the dashed compute-efficient frontier:

All three use synthetic placeholder data and are generated by assets/generate_showcase.py — run it to reproduce them.

Install

This repo is a Claude Code plugin marketplace. Add it, then install the plugin:

/plugin marketplace add Mapika/nice-figures
/plugin install nice-figures@nice-figures

To update later:

/plugin marketplace update nice-figures

Usage

Once installed, just describe the figure you want. Claude triggers the skill on requests like:

• "Make a training-curve plot of these RL scores with a smoothed trend and a shaded band, soft research-blog style."
• "Grouped bar chart comparing three models across four eval scenarios, with the rounded bar tops and error bars."
• "Scaling-law scatter on log-log axes with a power-law fit for our poster — clean white background."
• "Match the figure you made last week with the coral/peach bars."

You can also invoke it directly:

/nice-figures:nice-figures

Bring your own data (CSV or arrays) and Claude maps it onto the nearest recipe; describe a figure with no data and it generates a clearly-marked synthetic placeholder.

Other agents / plain Python

The style helper is a single dependency-light module, installable anywhere:

pip install git+https://github.com/Mapika/nice-figures
python -c "import soft_style"

AGENTS.md carries the agent-agnostic instructions (workflow, hard rules, the render-and-inspect loop) — point Codex/Cursor/your harness at the repo and it works without the Claude plugin machinery.

What's inside

AGENTS.md                        # agent-agnostic instructions (non-Claude harnesses)
plugins/nice-figures/skills/nice-figures/
├── SKILL.md                     # the skill: when to use it, conventions, gotchas
├── scripts/soft_style.py        # matplotlib style helpers (numpy + matplotlib only)
└── references/chart_recipes.md  # full code for all 16 chart archetypes
tests/                           # unit tests + every recipe executed end-to-end

The style

• Palettes — LINE_PALETTE (blue/mustard/sage/pink) for trend lines, BAR_PALETTE (coral/peach/gray/olive) for bars, MULTILINE_PALETTE for up to 5 categorical lines, plus sequential, diverging, and ordered-gradient colormaps.
• Rounded bars — the signature look. rounded_bars() / rounded_hbars() give softly rounded top corners computed in display space so they stay circular at any aspect ratio.
• Smoothing + bands — smooth_curve() and rolling_band() for trend lines with tight shaded uncertainty bands.
• Typography & axes — bold display title, gray subtitles and labels (never pure black), only bottom/left spines, no grid.
• Layout helpers — top_legend() (legend above the axes, never on the data), plain_log_ticks(), soft_colorbar().
• Paper mode — configure_style(scale=0.75) for single-column figures; Type 42 fonts embedded in every PDF.
• Export — save_figure() writes both PDF and PNG at 300 dpi.

The 16 archetypes

Trend-with-band, scatter + baseline, grouped bars, multi-line sweeps, heatmaps/confusion matrices, ROC/PR curves, distribution comparisons, box/violin, scaling-law fits, parity/calibration, 2D embeddings, ECDFs, forest plots, sorted horizontal bars, and Pareto fronts. Full copy-and-adapt code lives in references/chart_recipes.md.

Requirements

• matplotlib and numpy (the only hard dependencies of the style helper).
• The Inter font is preferred and picked up automatically if installed; otherwise it falls back to Helvetica → Arial → DejaVu Sans.

License

MIT