FigDataX

Figure Data eXtraction — high-precision scientific figure data extraction, as an agentic Claude Code skill.

Give Claude a figure image (or several) cropped from a paper; it extracts the numbers at up to ±0.5% accuracy and proves it with a validation overlay. The engine measures the geometry (plot area, tick pixel positions, series colors, sub-pixel centroids); Claude reads the semantics (tick values, legend) with vision; you are asked only when something is genuinely ambiguous.

Supported: scatter, line, bar (grouped/stacked), box, pie, heatmap, polar, error bars, log axes, multi-panel figures. Output: CSV per figure + optional multi-sheet Excel with provenance (method, calibration RMSE, validation rounds).

中文说明见 中文说明.md。

Install

git clone https://github.com/Shaowen-Ye/FigDataX ~/.claude/skills/FigDataX
bash ~/.claude/skills/FigDataX/scripts/setup.sh      # creates .venv, installs deps (uv)
~/.claude/skills/FigDataX/.venv/bin/python -m scripts.figdatax self-test

The folder name is FigDataX (case-sensitive on Linux).

Use

In Claude Code, just ask:

> 提取 /path/to/figure.png 图片数据
> Extract data from ./results/fig3.png
> 这几张图都提取一下，汇总成一个 Excel

Claude runs the autonomous loop in SKILL.md: classify the chart → engine geometry pass (figdatax geometry: bbox + tick positions + series colors) → visually verify the annotated overlay → pair tick values by vision → calibrate (hard RMSE < 1% gate) → extract sub-pixel centroids → visually check the validation overlay → iterate up to 3 rounds → deliver CSV/Excel + a provenance report.

Zero-typing entry points (macOS)

bash ~/.claude/skills/FigDataX/integrations/macos/install.sh

installs two shortcuts so you barely type at all:

• /figx slash command — in Claude Code, type /figx and drag the image(s) into the terminal. Full interactivity (Claude can still ask at the two gates).
• Finder Quick Action — select images in Finder → right-click → 快捷操作 → FigDataX 提取数据. Runs headless (claude -p, billed to your subscription), drops CSV/validation/Excel next to the images, then notifies you. In headless mode Claude cannot ask questions: at the gates it proceeds with best judgment and flags the uncertainty in the report. Model override: FIGX_MODEL=claude-opus-4-8.

Or use the CLI directly (see references/cli.md):

PY="$HOME/.claude/skills/FigDataX/.venv/bin/python"
$PY -m scripts.figdatax geometry figure.png --json geom.json --annotate geom.png
$PY -m scripts.figdatax extract figure.png \
    --calibration-points calib.json --color-target 0 255 255 --subpixel --validate

Documentation

• SKILL.md — the autonomous extraction loop (the single source of truth).
• references/ — precision & troubleshooting, morphological pipeline, special cases, full CLI reference.
• CHANGELOG.md — release notes.

History

A desktop app (PySide6) was developed and later discontinued in favor of the pure skill — Claude Code natively provides what the GUI required manual clicks for. Its source remains at tag app-v0.2.0.

License

MIT — see LICENSE.