Turning AI-generated paper figures back into editable Visio

What it does This is a Codex Skill (plus PowerShell scaffolding) that takes a reference image — a screenshot, a PNG, or an AI-generated model diagram — and rebuilds it as a Microsoft Visio .vsdx made of native shapes, editable text, connectors, and groups. From that single editable master it exports PNG, SVG, PDF, and PPTX. It also restyles existing Visio files against a reference and audits .vsdx packages for cheating.

The interesting bit The core rule is almost adversarial: the final file must not contain the reference image pasted in as a fake “reconstruction.” The skill explicitly forbids whole-image embedding and includes a package-inspection script to catch it. It also solves the classic multi-panel-figure failure mode — elements drifting across panel boundaries — with panel-local 0–1 coordinates and Assert-RelBox helpers that error out when a shape escapes its parent panel.

Key highlights

• Rebuilds model diagrams, architecture figures, and flowcharts as native Visio shapes, not bitmaps
• Single .vsdx master; SVG/PDF/PPTX/PNG are all exported from it, so formats stay consistent
• Panel calibration workflow with local-coordinate helpers (RectRel, TextRel, OvalRel, LineRel) to prevent overlap and drift in complex multi-panel figures
• Package inspection to detect .vsdx files that just embed the whole reference PNG
• Can restyle an existing Visio file (colors, fonts, layout) while keeping everything editable

Caveats

• Windows + Microsoft Visio + PowerPoint only — it drives everything through Visio COM Automation; draw.io users are told to adapt the interfaces themselves
• The author explicitly says it’s for model/architecture/flow diagrams; metric plots with real data are better done with scripts, not this
• Without Visio installed, you only get package checks and limited XML edits, not full reconstruction

Verdict Worth a look if you write papers in a Visio-based lab workflow and want an agent to produce figures you can actually nudge afterward. If you’re on Linux, use draw.io, or need data-driven plots, this isn’t your tool — and it knows it.