image-to-editable-ppt-skill: a Codex skill that rebuilds slide images as editable PowerPoint
Codex skill for converting slide images, PDFs, and image-based PPTX files into editable PowerPoint decks.
At a glance
- What is it?
- This Codex skill turns screenshots, PDFs and image-based PPTX decks into .pptx files where readable text becomes native text boxes and simple geometry becomes PowerPoint shapes. It is heavy, slow and token-hungry, and the README says so.
- Who is it for?
- Adopt it when you have a screenshot-style deck you must actually edit, you already run Codex with full access permissions, and you accept that a 10-page deck can consume a 5-hour quota. Do not adopt it to generate a new deck from an outline, or if you only want one slide touched up; the README points those cases at gpt-image-2.5-sunburst image editing and at codex-ppt-skill instead.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 3 days ago.
- What is it written in?
- Mainly Python, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 17, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
What image-to-editable-ppt-skill rebuilds, and for whom
The problem is the screenshot deck. Someone exports slides as images, or receives a PDF, or gets a .pptx where every page is one flattened picture. Nothing inside can be retyped, moved or restyled. This skill targets exactly that input: a single image, several images, a multi-page PDF, or an image-based PPTX, all reconstructed into an editable .pptx.
The reconstruction is partial by design. According to the README, readable text is restored as native text boxes where possible, simple geometry becomes PowerPoint shapes, and complex visual elements stay as separate image assets with a record of where they came from. The stated goal is that text, simple shapes and visual material can be adjusted separately afterwards.
It is not a deck generator. The README states plainly that the skill does not produce a new PPT from an article, report, outline or idea, and points that use case at a sibling project, codex-ppt-skill. The intended user is someone who has slides that exist only as pixels and needs to keep editing them.
How the pipeline normalizes pages, dispatches workers and restores text
Input is first normalized into per-page tasks, then rebuilt page by page. A single page or single image is handled by the main agent running the same page reconstruction flow locally. Multi-page input is dispatched to page workers or subagents and processed in parallel under a `max_concurrent_pages` setting. That means multi-page use depends on an environment where the agent can create subagents at all; the README says that if page workers cannot be created, you should move to an environment that supports them.
Text restoration is measurement-driven rather than guessed. During the prepare stage the skill generates a text annotation for each page containing box coordinates, font size and a font-size grouping. The model then restores text against those measured values, and the README states that text at the same level is automatically kept at a consistent size. If a Baidu AI Studio OCR token is configured, the skill calls PaddleOCR-VL to correct text boxes, font sizes and size groupings; without a token it degrades to built-in offline detection and the README warns that text restoration quality suffers.
Image work has a defined fallback chain. The agent calls Codex's built-in `image_gen.imagegen` first. Only when that tool is unavailable, errors, receives unreadable edit input, or returns no valid local image does the skill drop into the `editppt image` CLI, which tries local Codex OAuth (`~/.codex/auth.json`) and then an OpenAI-compatible API configured in `~/.editppt/config.yaml`. The README is explicit that missing parameters such as `mask`, `model`, `size`, `quality` or `out` do not by themselves trigger the fallback, and that the skill only accepts local paths the tool explicitly returns rather than scanning a directory for the newest file.
Two details are worth noting for anyone with a mixed deck. Multiple images become pages in the order supplied, while PDF and .pptx inputs keep their original page order. Speaker notes in a .pptx input are copied to the corresponding output page unchanged.
Installing the skill and running a first conversion
The skill lives in the `skills/` directory of the repository, and the README treats it as something you run inside Codex rather than a standalone CLI. The README recommends executing it with full access permissions, because it runs long and automatically performs OCR, image generation and editing, file reads and writes, subagent dispatch and long polling; approval modes interrupt those steps, and the README notes that auto-approve can still block at the OCR, image generation or third-party API stages.
If you use the CLI fallback path, third-party API credentials go in a config file under your home directory. On Windows the README gives the `%USERPROFILE%` form:
# ~/.editppt/config.yaml (Windows: %USERPROFILE%\.editppt\config.yaml)
# OpenAI-compatible API fallback used by `editppt image`The README does not print the full key list for that file, so read the repository docs before filling it in. The CLI surface it does document is small: `editppt image generate` and `editppt image edit`, where generation needs `--prompt` or `--prompt-file`, editing additionally needs `--image`, and page reconstruction should pass `--out` explicitly. Practical flags are limited to `--model`, `--size`, `--quality`, `--force`, `--dry-run`, `--timeout`, plus `--mask` for edits.
A first real run is therefore a prompt to Codex, not a shell command: point it at one slide image and ask it to rebuild that page as an editable .pptx. The README describes the expected result as a page where you can select and retype the text, drag simple shapes, and find the complex visuals as separate image assets. Expect the first page to take a while; the README says a single page can take more than 10 minutes.
Where this skill breaks down or is the wrong tool
Cost is the first constraint, and the README states it without hedging. Reconstructing a 10-page deck can exhaust a 5-hour quota, and the project recommends ChatGPT Pro users while telling Plus users to proceed cautiously. If you have no strong need for editability, the README advises against using the skill at all and suggests a lighter route: send the unsatisfactory slide image to gpt-image-2.5-sunburst image editing and get a modified image back.
Complex pages remain unpredictable. The README notes that complex pages can still consume a lot of tokens, and that when page validation fails the skill prefers local repair and reuse of already verified assets rather than regenerating everything for small imperfections. Asset extraction is not guaranteed faithful either: the README says a successful segmentation does not mean an icon matches the original, and items still need to be checked one by one.
Environment is a hard boundary. Multi-page conversion requires an agent that can dispatch page workers; without that capability the README tells you to switch environments. Non-Codex environments such as Claude Code, OpenClaw or Hermes Agent need an OpenAI-compatible API fallback unless Codex OAuth is available, which adds configuration the README says is usually unnecessary.
One more limit is easy to miss: the README states that reconstructed curve paths are not the same as data-driven charts. You can adjust a curve's shape and style in PowerPoint by right-clicking it, choosing the vertex editing option and dragging the white control handles, but the underlying numbers are not there.
How it differs from batch converters and from codex-ppt-skill
The natural alternative is an online batch conversion service, and the README names one directly in its sponsorship section: Codia NoteSlide, described as a fast, inexpensive online service suited to bulk conversion of many images or PDFs. The difference in approach is the point. A batch service maps pixels to a slide and returns a file quickly; this skill re-measures each page, separates text from shapes from imagery, and lets a model iterate on individual pages. The README frames the two as complementary, suggesting the skill for users who already subscribe to ChatGPT and want page-by-page reconstruction with repeated adjustment of text and layout.
The second alternative is the project's own sibling, codex-ppt-skill. They solve opposite problems. codex-ppt-skill generates a new deck from content; image-to-editable-ppt-skill reconstructs an existing visual page. The README points anyone whose goal is generating a deck to the former, and the repository includes a PDF, `assets/skill_duo_intro.pdf`, that introduces both.
If your input is a real, text-based PPTX rather than an image-based one, neither the batch route nor this skill is the right instrument. The skill's own scope is image-based PPTX input, and the README does not claim it repairs ordinary editable decks.
Maintenance, licence and what an upgrade costs you
The repository is not archived, and its last push was on 2026-09-16, one day before the date used for this review, so the codebase is being touched. The release history shows v0.4.0 on 2026-09-12, v0.3.3 on 2026-09-10 and v0.3.2 on 2026-07-19, which suggests a project that ships fixes in small increments rather than long release cycles.
The licence is MIT, which permits commercial use, modification and redistribution provided the copyright notice and permission notice are preserved. That is a statement about the licence text, not legal advice; if you redistribute the skill inside a product, have your own counsel read the LICENSE file in the repository root.
Upgrade cost is mostly operational, not code. Because the skill runs inside Codex and depends on built-in image tools, OCR and subagent dispatch, a new release can change how those steps are sequenced. The repository keeps a CHANGELOG.md at the top level, and that is the file to read before pulling a new version. The README also documents environment-specific paths, `~/.editppt/config.yaml` on Unix-like systems and `%USERPROFILE%\.editppt\config.yaml` on Windows, so any config you maintain lives outside the repository and survives a pull.
Editorial conclusion
Adopt it when you have a screenshot-style deck you must actually edit, you already run Codex with full access permissions, and you accept that a 10-page deck can consume a 5-hour quota. Do not adopt it to generate a new deck from an outline, or if you only want one slide touched up; the README points those cases at gpt-image-2.5-sunburst image editing and at codex-ppt-skill instead. Before committing, verify that your environment can dispatch page workers for multi-page input, and decide whether you will supply the Baidu AI Studio OCR token, since without it the README says text restoration falls back to offline detection at lower quality.
Frequently asked questions
How do I convert an image into an editable PPT with image-to-editable-ppt-skill?
Run the skill inside Codex with full access permissions, point it at the slide image, and let it normalize the page and rebuild it as a .pptx. The README recommends full access because the flow automatically performs OCR, image generation and editing, file reads and writes, and subagent dispatch.
How do I convert a non-editable PPT to an editable one?
If the PPTX is image-based, the skill accepts it as input and reconstructs each page, keeping the original page order and copying speaker notes to the corresponding output page unchanged. Readable text is restored as native text boxes where possible, simple geometry as PowerPoint shapes, and complex visuals stay as separate image assets.
Does image-to-editable-ppt-skill work with multi-page PDFs?
Yes. PDF input is normalized into per-page tasks and keeps its original page order. Multi-page input is dispatched to page workers and processed in parallel under `max_concurrent_pages`, so the environment must be able to create subagents.
Do I need an OCR token to use image-to-editable-ppt-skill?
The README says text size and position correction needs a free Baidu AI Studio OCR token, which the skill uses through PaddleOCR-VL. Without it, the skill falls back to built-in offline detection and the README warns that text restoration quality is reduced.
How long does one page take to convert with image-to-editable-ppt-skill?
The README states that a single page can take more than 10 minutes to reconstruct, and that a 10-page deck may exhaust a 5-hour quota. It recommends ChatGPT Pro and advises Plus users to be cautious.
Community notes