Skill: Visualization

Use this skill whenever the user asks to:

• Plot, visualize, chart, or graph data.
• Create figures for EDA, reports, dashboards, or presentations.
• Modify or improve existing matplotlib / seaborn code.
• Save figures to disk for reports or notebooks.

Library

• Use matplotlib for charts.
• Use seaborn alongside matplotlib for theming, palettes, and statistical plots.

Global Setup (do this first)

At the start of any notebook or script that produces figures, set the theme once. font_scale enlarges all text elements proportionally, so individual fontsize= arguments are usually unnecessary.

import matplotlib.pyplot as plt
import seaborn as sns

sns.set_theme(
    style="whitegrid",
    palette="muted",
    font_scale=1.2,
)

Japanese text

If any text (titles, labels, legends, annotations) contains Japanese, configure a CJK-capable font, otherwise characters render as tofu (□□□).

Preferred approach (cross-platform):

import japanize_matplotlib  # uv add japanize-matplotlib

Alternative (set an installed CJK font explicitly):

plt.rcParams["font.family"] = "Noto Sans CJK JP"  # or "IPAexGothic", "Hiragino Sans", "Yu Gothic"

Figure Creation

• Do not use the stateful plt.figure(...) / plt.plot(...) style.
• Always create figures and axes explicitly, and prefer constrained_layout=True over tight_layout():

fig, ax = plt.subplots(figsize=(10, 6), constrained_layout=True)

• Use the object-oriented API: ax.set_title(), ax.set_xlabel(), ax.plot(), etc.
• figsize=(10, 6) is a reasonable default. Adjust to content: wide time series → (12, 4), square scatter → (6, 6), multi-panel → scale up accordingly.

Color Map / Color Palette

Choose by data type:

• Categorical: "muted", "Set2", "colorblind" (seaborn palettes).
• Sequential (continuous): "viridis", "cividis", "mako" — perceptually uniform.
• Diverging: "coolwarm", "RdBu", "vlag".
• Forbidden: "jet", "rainbow" — not perceptually uniform, poor for colorblind viewers.

Font Size

sns.set_theme(font_scale=1.2) covers most cases. Override per-element only when needed (e.g. a long title needs to be smaller, or one label needs emphasis):

ax.set_title("...", fontsize=18)

Do not repeat fontsize= on every call — it is redundant when font_scale is set.

Axis Scale Guidelines

• Bar charts: start y-axis at 0 (ax.set_ylim(bottom=0)). Truncated bars are misleading.
• Line / scatter: do not force y-axis to 0 — it can hide meaningful variation. Let matplotlib autoscale, or set limits based on the data range.
• Log scale: use ax.set_yscale("log") when data spans multiple orders of magnitude.

Legend

• Use ax.legend() only when ≥2 series are plotted.
• If the auto-placement overlaps data, set explicitly: ax.legend(loc="upper left") or place outside: ax.legend(loc="center left", bbox_to_anchor=(1.0, 0.5)).

Saving

• Save final figures under outputs/figures/.
• Diagnostic figures produced by a *-diagnostics skill go under outputs/diagnostics/<YYYYMMDD-HHMM>_<skill>/ instead (gitignored; the report links the directory and shows only the 判定表).
• Use the path utilities from src/analysis_project/paths.py.

from analysis_project.paths import outputs_dir, ensure_parent_dir

# 出力パスを構成して保存
output_path = outputs_dir() / "figures" / "survival_by_class.png"
ensure_parent_dir(output_path)
fig.savefig(output_path, dpi=150)

• Use descriptive snake_case filenames: monthly_sales_2024h1.png, not fig1.png.
• DPI guidance:
  • Notebook / README / slides: dpi=150
  • Publication / print: dpi=300
• Always call plt.close(fig) after saving to free memory.

fig.savefig("outputs/figures/monthly_sales.png", dpi=150, bbox_inches="tight")
plt.close(fig)

Chart Quality Checklist

Before finalizing a chart, verify:

• Title clearly describes what the chart shows.
• Axis labels include units where applicable (e.g. "売上 (万円)", "Latency (ms)").
• Font sizes are readable (rely on font_scale=1.2 as baseline).
• Color palette is perceptually uniform / colorblind-friendly (no jet/rainbow).
• Date range noted in title, subtitle, or annotation when relevant.
• Sample / filter note when data is subsetted (e.g. "n=1,234, 2024年1月〜6月").
• Bar charts start y-axis at 0; other chart types use sensible limits.
• Legend present when multiple series; placement does not overlap data.
• Japanese text renders correctly (japanize-matplotlib or CJK font configured).
• Saved to outputs/figures/ with descriptive filename, followed by plt.close(fig).

Example

from pathlib import Path
import matplotlib.pyplot as plt
import seaborn as sns
import japanize_matplotlib  # noqa: F401  # 日本語フォント有効化

sns.set_theme(style="whitegrid", palette="muted", font_scale=1.2)

Path("outputs/figures").mkdir(parents=True, exist_ok=True)

fig, ax = plt.subplots(figsize=(10, 6), constrained_layout=True)

ax.bar(categories, values)
ax.set_title("月別売上推移 (2024年1月〜6月)")
ax.set_xlabel("月")
ax.set_ylabel("売上 (万円)")
ax.set_ylim(bottom=0)  # 棒グラフは0起点

fig.savefig("outputs/figures/monthly_sales_2024h1.png", dpi=150, bbox_inches="tight")
plt.close(fig)