Figure agent

<SUBAGENT-STOP> If you were dispatched as a subagent to generate a specific figure from data or create a specific diagram, skip this skill. Do the task and return structured results immediately. </SUBAGENT-STOP> <HARD-GATE> NEVER fabricate figure data. Every figure must be backed by real experimental results or a traceable data source. If no data is provided, ask for it before generating anything. </HARD-GATE>

Figure Agent Skill

Decision-driven figure generation for conference papers. Analyzes the paper content to determine figure types, generates Matplotlib/Seaborn code for quantitative figures, and iterates on quality via critic feedback.

Figure Type Decision Agent

When generating figures for a paper, first analyze the content to determine the appropriate figure type:

Quantitative Figures (Code-generated)

Use when the figure shows actual experimental results, metrics, or data:

Figure Type	Best For	Matplotlib Function
Bar chart	Comparing discrete conditions, ablations	plt.bar()
Line plot	Convergence curves, training trajectories	plt.plot()
Scatter plot	Correlation analysis, parameter sensitivity	plt.scatter()
Heatmap	Ablation matrices, hyperparameter grids	plt.imshow() or sns.heatmap()
Box plot / Violin	Distribution of results across seeds	plt.boxplot() or sns.violinplot()
Histogram	Result distributions	plt.hist()
Stacked bar	Comparison across multiple metrics	plt.bar(stacked=True)
Grouped bar	Multiple methods across conditions	plt.bar() with offset positions

Architecture/Conceptual Figures (Diagram)

Use when the figure shows the proposed method, system architecture, or conceptual framework. These should be drawn with Excalidraw or a similar diagramming tool. For code-generated alternatives, use Matplotlib patches.

Selection Criteria

Is the figure based on experimental data?
├── YES → Is it a comparison (method A vs B) or relationship (parameter vs metric)?
│   ├── Bar/Line/Scatter/Box → Generate Matplotlib code
│   └── Architecture/Flowchart → Generate diagram code
└── NO → It must be conceptual or illustrative
    └── Architecture/Flowchart → Generate diagram code

Matplotlib/Seaborn Code Generation

Template

import matplotlib.pyplot as plt
import matplotlib as mpl
import numpy as np
import json

# === LOAD DATA ===
with open("results.json") as f:
    data = json.load(f)

# === STYLE SETUP ===
plt.rcParams.update({
    "figure.figsize": (5.0, 3.5),  # NeurIPS column width
    "font.family": "serif",
    "font.size": 10,
    "axes.titlesize": 11,
    "axes.labelsize": 10,
    "xtick.labelsize": 9,
    "ytick.labelsize": 9,
    "legend.fontsize": 9,
    "lines.linewidth": 1.5,
    "axes.linewidth": 0.8,
    "grid.alpha": 0.3,
    "savefig.dpi": 300,
    "savefig.bbox": "tight",
})

# === PLOT COLOURS (Color-blind friendly) ===
CB_COLORS = ["#0072B2", "#E69F00", "#009E73", "#CC79A7", "#56B4E9", "#F0E442"]

# === FIGURE ===
fig, ax = plt.subplots(1, 1)

# ... plotting code ...

# === STYLE ===
ax.set_xlabel("X Label")
ax.set_ylabel("Y Label")
ax.set_title("Title")
ax.legend()
ax.grid(True, alpha=0.3)

# === SAVE ===
plt.tight_layout()
plt.savefig("figure_1.pdf", dpi=300)
plt.savefig("figure_1.png", dpi=300)
plt.close()
print("FIGURE_SAVED: figure_1.pdf")

Conference Style Presets

NeurIPS (default)

plt.rcParams.update({
    "figure.figsize": (5.0, 3.5),   # Single column
    "font.family": "serif",
    "font.size": 10,
})

ICML

plt.rcParams.update({
    "figure.figsize": (5.5, 3.5),   # Slightly wider
    "font.family": "serif",
    "font.size": 9,
})

ICLR

plt.rcParams.update({
    "figure.figsize": (5.0, 3.5),
    "font.family": "serif",
    "font.size": 10,
})

Color Schemes

Use color-blind friendly palettes by default:

# Qualitative (for categories)
CB_COLORS = ["#0072B2", "#E69F00", "#009E73", "#CC79A7",
             "#56B4E9", "#F0E442", "#D55E00", "#000000"]

# Sequential (for heatmaps)
CB_SEQUENTIAL = ["#FFFFFF", "#F0E442", "#D55E00", "#000000"]

# Diverging (for difference plots)
CB_DIVERGING = ["#0072B2", "#FFFFFF", "#D55E00"]

Iterative Improvement Loop

For each figure, run a 3-step improvement loop:

Step 1 — Generate

Generate the initial figure code from the data and figure specification.

Step 2 — Critic

Analyze the generated figure:

• Is the data visualization accurate? No misleading axis scales?
• Is the figure readable at conference page width (~5 inches)?
• Are labels clear and font sizes appropriate?
• Is the color scheme accessible (color-blind friendly)?
• Is the caption informative (takeaway in one sentence)?
• Does the figure serve the paper's core claim?

Step 3 — Improve

Apply critic feedback. Iterate up to 3 times or until the critic judges the figure as "publishable."

Figure Specification Format

Each figure must be specified with:

figure:
  id: 1
  type: "bar"           # bar | line | scatter | heatmap | box | histogram
  caption: "Comparison of convergence rates across optimizers. SGD-Adam achieves 15% faster convergence than baselines."
  data_source: "results.json"  # or inline data
  x: "optimizer"        # column or variable name
  y: "convergence_time" # column or variable name
  hue: "dataset"        # optional grouping variable
  style: "NeurIPS"      # NeurIPS | ICML | ICLR
  width: 5.0            # inches
  height: 3.5
  color_scheme: "colorblind"

Output Structure

Figures are saved to sessions/<slug>/figures/:

figures/
├── figure_1.pdf      # Vector format for paper
├── figure_1.png      # Raster format for preview
├── figure_1_caption.md  # Caption text
├── figure_1_code.py  # Regeneratable code
├── figure_2.pdf
├── figure_2.png
├── figure_2_caption.md
└── figure_2_code.py

Anti-patterns

• Jupyter notebook output as a figure. Conference papers need vector graphics (.pdf). Never submit screenshots or Jupyter inline plots.
• Rainbow color schemes. Use color-blind friendly palettes. Never use jet/rainbow colormaps for quantitative data.
• 3D plots for 2D data. 3D plots rarely add value and often obscure the data. Use 2D alternatives unless 3D is essential.
• Figures without captions. Every figure must have a caption with a takeaway. If there is no takeaway, reconsider whether the figure belongs.
• Axis manipulation. Never truncate the y-axis to exaggerate differences. Always start at 0 for bar charts unless there is a principled reason.

Verification Checklist

• Figure type matches the data type (quantitative vs conceptual)
• All data values are real (not fabricated)
• Color-blind friendly palette is used
• Font sizes are readable at conference page width
• Caption includes a one-sentence takeaway
• Figure is saved as PDF (vector) and PNG (raster)
• Code is saved for regeneration
• Critic judged the figure as "publishable"
• Axis scales are not misleading