scibox-diagram

Copy the template's example.json from the assets directory, fill in your content with manual line breaks, then run the matching render script such as python3 scripts/roadmap_5band.py content.json -o out.drawio. The script validates every slot's character budget before writing the file.

Run python3 scripts/export_figure.py fig.drawio to produce a 1:1 PNG and a vector PDF. It requires the drawio desktop command line (brew install --cask drawio on macOS); without it, use preview_html.py for a browser preview instead.

It covers schematic diagrams: five-band technical roadmaps, three-column research frameworks, staged process flows, landscape task pipelines, algorithm flowcharts, and model architecture diagrams. Data charts like line plots and heatmaps are out of scope and belong to the scibox-figure skill.

draw.io's automatic wrapping handles Chinese poorly, so text must be manually broken with \n and sized by the rule: full-width characters equal the font size in pixels, half-width characters half that, line height about font size plus 3. The check_layout.py script detects overflow before rendering.

Yes, the replication workflow calibrates coordinates, colors, and font sizes from the reference image using connected-component analysis, then iterates at least three rounds of render-compare-fix with pixel differencing. Approximated elements are recorded in an asset ledger rather than silently dropped.

Avoid it for statistical data charts such as line plots and heatmaps, which belong to scibox-figure, and for LaTeX formula derivation chains, which are better done in TikZ or inline text. It targets editable schematic diagrams, not data visualization.