A geologist’s AI illustrator that refuses to flatten your figures

What it does

Scientific Illustrator is a Codex/MCP plugin that takes a reference image and reconstructs it as editable objects—text, shapes, arrows, tables, and charts—inside Microsoft PowerPoint, WPS Presentation, or draw.io. Instead of flattening everything into a static screenshot, it redraws the figure so fonts, arrows, and axes remain tweakable later. When it hits a micrograph or complex texture it genuinely cannot vectorize, it crops only that minimal patch to a bitmap and keeps the rest editable.

The interesting bit

The author—who identifies only as “a geology PhD”—built a Designer-Drawer-Reviewer-Corrector loop into the prompt pipeline, forcing the model to verify structure, preview screenshots, and fix errors region-by-region before declaring success. It also behaves like a polite coworker: by default it draws in the background without seizing window focus, and if draw.io hands it an unknown shape name, it throws an explicit error rather than silently degrading to a rectangle.

Key highlights

• Outputs native editable formats: PPTX for PowerPoint/WPS, .drawio for diagrams.
• Cross-platform support for Windows and macOS, with optional real-time per-object rendering in Mac PowerPoint via an Office.js add-in.
• Defensive validation: verifies document open states, distinguishes WPS from PowerPoint, and validates shape names on the draw.io canvas.
• CI runs on Ubuntu, macOS, and Windows, though the README notes GitHub’s public runners lack commercial Office licenses, so live Windows app integration relies on post-install state checks.

Caveats

• The optional Mac PowerPoint real-time mode requires manual certificate trust and sideloading an Office.js add-in; the plugin will not auto-trust system certificates.
• Because GitHub Actions runners do not include licensed PowerPoint or WPS, automated Windows in-app integration testing is limited to state-tool verification rather than full live-app validation.

Verdict Worth a look if you routinely need publication-style figures that collaborators or journals will later edit in PowerPoint or draw.io. If you just need a one-off raster export, it is probably overkill.