image-to-editable-ppt-skill

This is a skill for Codex (an AI coding assistant from OpenAI) that converts slide images, multi-page PDFs, and image-based PowerPoint files into editable .pptx files. The goal is to take a presentation that exists only as images and recover as much editable structure as possible: text becomes real text boxes you can edit, simple shapes become native PowerPoint shapes, and complex visual elements are saved as separate image assets.

The skill uses a multi-agent architecture. A main agent breaks the input into individual pages and assigns each page to its own sub-agent. These page sub-agents work in parallel to reconstruct the slides, which speeds up processing for presentations with many pages. The main agent then checks the quality of each page, schedules repairs if needed, and assembles the final .pptx file. No third-party OCR service is required; the reconstruction is done visually using the models available within Codex.

The supported inputs are a single image, a collection of multiple images, a multi-page PDF, or an image-based PowerPoint file. In all cases the output is a .pptx file with the original page order preserved. If the input was a .pptx file, speaker notes from the original slides are copied over unchanged.

For complex visual elements such as photographs, illustrations, or decorative graphics, the tool saves them as independent image assets rather than attempting to recreate them as editable shapes. Fully pixel-perfect reproduction is not guaranteed; some text and elements may shift slightly in position.

This skill is designed specifically for Codex and has been tested with Codex Plus and Max memberships. It is not intended for generating new PowerPoint content from scratch. The project is written in Python, documented primarily in Chinese with an English README also available, and is licensed under MIT.