Ds Figure Polish

Figure Polish

Use this skill when a figure matters beyond transient debugging.

This includes:

• a main-experiment summary image sent to a connector
• an aggregated analysis-campaign chart
• a paper-facing main figure
• an appendix / supplementary figure
• any figure that will be stored as a durable artifact or cited in writing

Do not use this skill for disposable debug plots unless the user explicitly asks for them to be polished.

Core principle

DeepScientist figures should feel academic, restrained, and clear.

The goal is not to make a plot “fancy”. The goal is to make the intended comparison obvious without visual clutter.

Use one dominant message per figure. If multiple unrelated claims are competing inside the same image, split the figure instead of cramming everything into one panel.

Surface classes

First classify the figure:

• connector_milestone
  • quick summary image for QQ / chat / copilot milestone reporting
  • usually png
  • message-first and minimal
• paper_main
  • core paper figure
  • export pdf or svg plus a png preview
  • must remain readable after likely single-column or double-column placement
• appendix
  • supplementary figure
  • may contain slightly more detail, but still avoid dashboard clutter
• internal_review
  • used for local diagnosis and internal comparison
  • can be lighter-weight, but still should follow the same visual discipline if it may later be promoted

Style contract

Prefer the bundled Matplotlib style asset when plotting in Python:

• assets/deepscientist-academic.mplstyle

If you need a custom script, start from that style instead of inventing a fresh bright theme.

Default visual rules:

• white or near-white background
• muted Morandi palette only
• no neon colors
• no rainbow / jet-like colormaps
• no heavy shadows, glossy gradients, or thick black borders
• top and right spines removed unless a special plot truly needs them
• light grid only when it helps reading values
• legend minimal; prefer direct labeling when it is clearer
• main method should be visually dominant
• baseline or comparison lines should be slightly more neutral than the main method

Chart selection

Choose the chart by the research question:

• line chart
  • trends over steps, epochs, budgets, or ordered scales
• bar chart
  • a small number of categorical end-point comparisons with a meaningful zero baseline
• point-range / dot plot
  • comparisons where uncertainty, confidence intervals, or seed spread matter
• box / violin / histogram
  • only for true distribution questions with enough samples
• heatmap
  • only when the matrix structure itself is the result

Do not use heatmaps or crowded dashboards just because they look “richer”.

Continuous color rules

• ordered magnitude -> sequential muted palette
• signed delta around zero or a reference -> diverging muted palette with a neutral midpoint
• categories -> discrete palette only

Avoid any colormap whose lightness jumps back and forth or whose hue changes overwhelm numeric ordering.

Mandatory render-inspect-revise workflow

If a figure is intended for milestone reporting, paper drafting, appendix use, or durable artifact storage, you must follow this sequence:

1. render a first draft
2. open the rendered figure yourself with the available file / image inspection capability
3. inspect the actual result, not just the plotting code
4. revise the figure if readability or composition is weak
5. re-export the final version

Do not treat a figure as final if you have not inspected the rendered result. Do not assume “the code looks fine” means “the figure looks fine”.