image-to-editable-ppt-skill: Rebuild Slide Images as Editable PowerPoint

Someone sends you a gorgeous deck — as screenshots. Or a PDF slideshow, or an old PPTX exported as images. The content is all there, but every word is pixels. ningzimu/image-to-editable-ppt-skill rebuilds those pages as an editable .pptx: readable text comes back as native text boxes, simple geometry becomes real PowerPoint shapes, and complex visuals stay as separate image assets with source records.

Why This Skill Matters

The differentiator is measurement, not eyeballing. During preparation, each page gets a text annotation pass — box coordinates, font sizes measured from the source image, and same-size text grouped together — so the model restores type against measured values instead of guesses. The optional OCR correction layer uses Baidu's PaddleOCR-VL through a free AI Studio token, and without a token the skill degrades to built-in geometric detection with lower text fidelity.

This is a heavyweight, multi-agent rebuild rather than a lightweight converter. Single-page inputs are handled by the main agent locally; multi-page decks get dispatched to page workers that run in parallel up to max_concurrent_pages, each doing a rebuild-then-self-check-then-correct loop until the page is close to the original.

Installation

The README documents a natural-language install — you give the agent the repo URL and it sets the skill up, including the editppt CLI it depends on:

Install the image-to-editable-ppt skill from
https://github.com/ningzimu/image-to-editable-ppt-skill

In Codex, invoke it explicitly with $image-to-editable-ppt. The skill needs an agent that supports skill loading, file read/write, and CLI execution; multi-page work additionally needs subagent dispatch.

Real Workflow: A Screenshot Deck Back Into Rotation

1. Attach the slide images, the PDF, or the image-based PPTX to the conversation, then prompt:

$image-to-editable-ppt Convert this deck into an editable PPT.

2. The skill normalizes input into per-page tasks, rebuilds each page, and assembles the final .pptx in page order. Input from a .pptx keeps one promise worth knowing: original speaker notes are copied to the output unchanged.
3. Open the result in PowerPoint and edit text directly — same-size text lines come back grouped, so font sizes stay consistent automatically.
4. Compare the source image against the output page to catch missing text, misplacements, or assets the validator flagged.

Tips

• Apply for the free Baidu AI Studio token before your first run — the AI asks for it once and writes it into user-level config; text fidelity is noticeably better with it.
• Run it in Codex with full-access mode enabled; the README warns that approval prompts interrupt the OCR and image-generation stages and can stall the pipeline.
• Image generation prefers Codex's built-in image_gen.imagegen, falling back to Codex OAuth, then to an OpenAI-compatible API configured in ~/.editppt/config.yaml.
• Budget honestly: the README warns a 10-page deck can consume your full 5-hour usage window, and single pages can take over 10 minutes.

When Not to Use This

The README is blunt about the trade: rebuilding can cost 2-3x what generating an image-based deck costs, and fidelity below gpt-5.5-class models is not guaranteed. If you just want one slide tweaked, send that image to an image editor instead. And if your actual goal is a new deck from an outline, that is the companion codex-ppt-skill — this skill rebuilds existing pages; it does not write new content.

See the leaderboard for more skills.