Matplotlib render review

Matplotlib Render-Review-Refine Workflow

Claude Code is multimodal -- it can see rendered PNGs via the Read tool. Use this to close a visual feedback loop: generate chart code, render to PNG, Read the PNG, review what you see, and iterate.

---

Section 1: Render-Review-Refine Workflow

Every chart task follows one of two workflows. Both end with visual verification.

Workflow A -- New Chart from Scratch

1. Write chart code as a standalone .py script that:
   • Reads the CHART_OUTPUT env var for the output path (os.environ.get("CHART_OUTPUT", "/tmp/chart_review.png"))
   • Uses OOP Figure API: Figure() + FigureCanvasAgg(fig) + fig.savefig()
   • Constructs minimal test data inline OR imports from the project
2. Render via the harness:

   python ${CLAUDE_SKILL_DIR}/scripts/render_review.py /tmp/my_chart.py --output /tmp/my_chart.png
   

3. Read the PNG with the Read tool to visually inspect the result.
4. Apply the Self-Review Checklist (Section 2).
5. Fix and re-render if issues found (max 2 additional cycles).
6. Integrate the chart code into the target module. Clean up temp files.

Workflow B -- Modify Existing Chart Code

1. Read the existing rendering function in the target module.
2. Make the change directly in the target file.
3. Write a minimal test harness script to /tmp/ that imports the modified code and renders a representative case with realistic data.
4. Render via the harness, Read the PNG, apply checklist.
5. Fix and re-render if issues found (max 2 additional cycles).
6. Clean up temp files.

Workflow Rules

• Maximum 3 total render cycles (initial + 2 refinements). If issues remain after 3 cycles, report what is still wrong to the user and stop.
• Always clean up /tmp/chart_*.py and /tmp/chart_*.png files when done.
• DPI: 150 for all renders (good balance of detail and file size). Lower DPI loses review detail.
• The render harness passes CHART_OUTPUT env var to the script -- always read it with os.environ.get("CHART_OUTPUT", "/tmp/chart_review.png").

---

Section 2: Self-Review Checklist

After every Read of a rendered PNG, check these items before deciding whether to iterate or accept.

Layout and Spacing

• Panels are not overlapping or clipping into each other
• Title is visible and not cut off by the figure boundary
• All axis labels are fully readable (not truncated or overlapping)
• Legend (if present) does not obscure data
• Sufficient padding between subplots (no label collisions)

Data Correctness -- Line Plots

• Lines track the expected trend (increasing, decreasing, or flat as expected)
• No unexpected gaps or discontinuities in the line
• Multiple lines are distinguishable from each other
• Data points fall within the expected value range

Data Correctness -- Bar Charts

• Bar heights are proportional to the underlying data values
• Bars are not clipped at the top or bottom of the axes
• Grouped/stacked bars align correctly with their categories
• Bar labels (if present) match their corresponding values

Data Correctness -- Scatter Plots

• Points appear at the correct (x, y) positions
• Point sizes and colors (if mapped to data) vary as expected
• No points are hidden behind others without transparency or jitter
• Outliers are visible and not clipped by axis limits

Data Correctness -- Heatmaps

• Color gradient maps to values in the correct direction
• NaN or missing cells are visually distinct (transparent or hatched)
• Row and column labels are readable, not squeezed or overlapping
• Colorbar is present and labeled with units

Color and Contrast

• Distinct data series use clearly distinguishable colors
• Grid lines are subtle (alpha ~0.3), not competing with data
• Text is readable against the background
• Color choices work for common forms of color blindness (avoid red/green only)

Axes

• Axis labels describe the data with units where appropriate
• Tick labels are formatted for readability (K/M/B suffixes for large numbers, appropriate date formats for time series)
• No axis is entirely blank when data was expected
• Shared axes are properly aligned across panels

---

Section 3: Figure Creation Quick Reference

The 3-line idiom for every chart (OOP API, no pyplot):

from matplotlib.figure import Figure
from matplotlib.backends.backend_agg import FigureCanvasAgg

fig: Figure = Figure(figsize=(12, 8))
FigureCanvasAgg(fig)  # Required for fig.savefig() and _repr_png_
# ... add axes, draw data ...
fig.savefig(path, dpi=150, bbox_inches="tight")

Standard DPI: 150 for both production and review renders.

For multi-panel layouts, use GridSpec:

import matplotlib.gridspec as gridspec

gs = gridspec.GridSpec(n_rows, 1, figure=fig, height_ratios=ratios, hspace=0.35)
ax_top = fig.add_subplot(gs[0, 0])
ax_bottom = fig.add_subplot(gs[1, 0], sharex=ax_top)

---

Section 4: Render Harness Usage

The render harness script lives at ${CLAUDE_SKILL_DIR}/scripts/render_review.py.

# Basic render (output defaults to /tmp/chart_review.png)
python ${CLAUDE_SKILL_DIR}/scripts/render_review.py /tmp/my_chart.py

# Custom output path
python ${CLAUDE_SKILL_DIR}/scripts/render_review.py /tmp/my_chart.py --output /tmp/custom.png

# Render and auto-cleanup the script file
python ${CLAUDE_SKILL_DIR}/scripts/render_review.py /tmp/my_chart.py --cleanup

The harness:

• Passes CHART_OUTPUT env var to the script with the resolved output path
• Prints the absolute PNG path on success (use this path with the Read tool)
• Prints diagnostic stderr on failure
• Returns non-zero exit code on any error

Your chart script should read the output path from the environment:

import os
output_path = os.environ.get("CHART_OUTPUT", "/tmp/chart_review.png")
fig.savefig(output_path, dpi=150, bbox_inches="tight")