FigureClaw

English | 简体中文

Make scientific figure creation easy: FigureClaw is designed for high-quality research visualization, automatically recommending the most suitable chart and generating reproducible Python plotting code in one click—helping you create publication-ready figures effortlessly.

Executable-first scientific figure skills with runnable Python plotting code for Codex, Claude Code, and Dr. Claw.

Why FigureClaw

In the era of scientific automation, the biggest pain point is still: it's hard to make beautiful, expressive, and publication-ready figures.

• The difference between top-tier and ordinary journals often lies in the quality and clarity of figures.
• Most researchers and developers struggle to create high-quality, reproducible visualizations, limiting the impact of their work.
• FigureClaw makes "top-journal-level" figures easy, automatic, and reproducible—helping your results stand out.

With FigureClaw, you get:

• One unified setup prompt—no more long manual install guides
• Only charts with ready-to-run local code templates are recommended by default
• Style-heavy or unsupported charts (like sunburst, chord) are exposed as conceptual options, not fake code
• Examples, audits, packaging tools, and local references—all in one portable skill package

60-Second Quick Start

FigureClaw is Codex-first for first-time setup.

Ask your agent to read the unified setup guide:

Read https://raw.githubusercontent.com/Boom5426/FigureClaw/refs/heads/main/setup.md and set up FigureClaw for me.

Then verify the install from the repository root:

python3 skills/figure-recommender/scripts/generate_figure_response.py \
  --brief-file skills/figure-recommender/examples/briefs/grouped-comparison.json \
  --output json

You will get:

• an executable primary_chart
• "contrast_dot" as the default chart id for the grouped comparison example
• a palette recommendation
• dependency hints
• runnable Python plotting code

Manual install for Codex

git clone https://github.com/Boom5426/FigureClaw.git ~/.codex/FigureClaw
mkdir -p ~/.codex/skills
ln -sfn ~/.codex/FigureClaw/skills/figure-recommender ~/.codex/skills/figure-recommender

Restart Codex after the link is created, then re-run the smoke test from ~/.codex/FigureClaw.

What You Can Generate

FigureClaw currently ships executable templates for:

• group comparison figures
• composition figures
• distribution figures
• line and multi-series trend figures
• matrix heatmaps
• benchmark scatter plots with error bars
• weighted relation and network-style figures

Conceptual requests such as sunburst or chord are still supported, but they are exposed as conceptual choices while FigureClaw emits executable code from a supported chart.

Install

The preferred entrypoint is setup.md, which routes the user to the right flow for the current environment and includes one shared smoke test.

Install With Codex

Codex-first users should start here.

Paste this into Codex:

Fetch and follow instructions from https://raw.githubusercontent.com/Boom5426/FigureClaw/refs/heads/main/.codex/INSTALL.md

Manual path target:

~/.codex/skills/figure-recommender

Success looks like:

• the skill exists at ~/.codex/skills/figure-recommender
• the smoke test prints "primary_chart"
• the grouped comparison example resolves to "contrast_dot"
• the response contains "python_code"

Install With Claude

Paste this into Claude Code:

Fetch and follow instructions from https://raw.githubusercontent.com/Boom5426/FigureClaw/refs/heads/main/.claude/INSTALL.md

Manual path target:

~/.claude/skills/figure-recommender

Install With Dr. Claw

1. Run python3 skills/figure-recommender/scripts/package_skill.py
2. Upload the generated dist/figure-recommender.zip in the Dr. Claw Skills UI
3. Let Dr. Claw discover the packaged SKILL.md, templates, references, and examples

Verify

Run the shared smoke test from the repository root:

python3 skills/figure-recommender/scripts/generate_figure_response.py \
  --brief-file skills/figure-recommender/examples/briefs/grouped-comparison.json \
  --output json

Confirm the response includes:

• "primary_chart"
• "contrast_dot"
• "python_code"

Example Workflow

The default workflow is:

1. describe the figure goal
2. resolve or infer a structured figure_brief
3. choose the best executable chart
4. choose a compatible palette
5. emit runnable Python code from local templates

Minimal brief shape:

{
  "story_goal": "compare_group_difference",
  "field_mapping": {
    "category": "condition",
    "value": "score"
  }
}

Defaults:

• figure_role = paper-main
• style_mode = readable
• palette_mode = auto

Starter examples live under skills/figure-recommender/examples/briefs/.

How It Works

FigureClaw follows an executable-first contract:

1. Normalize and validate one brief object
2. Rank chart candidates
3. Pick the best executable primary_chart
4. Optionally expose a conceptual_chart when the user explicitly asks for an unsupported chart family
5. Render Python code from the shipped local template set

The result always keeps the executable chart and generated code aligned.

Repository Structure

• setup.md: unified agent-readable setup entrypoint
• skills/figure-recommender/: runtime package, references, templates, and examples
• .codex/INSTALL.md: Codex-specific install flow
• .claude/INSTALL.md: Claude Code-specific install flow
• docs/source-audits/: notebook source audit artifacts
• tests/: regression, packaging, validation, and selection tests

Manual install

Use the collapsed Codex block above for first-time manual setup. Keep setup.md as the primary entrypoint unless you are debugging a broken local installation.

Development

Run the full test suite:

python3 -m pytest tests -q

Rebuild the zip package:

python3 skills/figure-recommender/scripts/package_skill.py

Regenerate notebook audit artifacts:

python3 skills/figure-recommender/scripts/export_source_notebooks.py \
  --output-dir docs/source-audits