GordenSuperPPTSkills: three Codex skills that turn generated slide images into editable PPTX

What GordenSuperPPTSkills solves, and who it is actually for

Most AI slide tools produce either a deck you cannot edit or an editable deck that looks like a template. This project takes a different route: let GPT generate each slide as a full image, then run a second pass that decomposes the image back into editable PowerPoint objects. The README frames the goal as "使用GPT生成豪华的图片格式PPT，然后转换为完全可编辑的PPTX文件". The audience is narrow and stated plainly: "仅限Codex使用", Codex only. If you do not run Codex, the skills have no supported host. The README notes that Opus plus a GPT image API could in theory do the same work, but that no adaptation for that path exists. So the real user is someone with a Codex environment, a GPT 5.5 model, and a tolerance for spending subscription quota on image conversion.

Three skills, one pipeline: how the image-to-PPTX decomposition works

The repository is not one skill but three, each self-contained with its own scripts/ and references/ directories. GordenImagePPTGen takes a topic or content brief and produces one .png per slide plus an image-based .pptx. GordenImage2PPTX takes images and reconstructs an editable .pptx. GordenSuperPPTSkill chains the two in sequence. The mechanism behind the second skill is a layer split. The README describes extracting the background image, the frame image, icons and decorative elements, and text, then reassembling them in PowerPoint by coordinate. That is why the conversion is not a simple vectorization: it is a vision-driven segmentation task, and the README says many detail checks and constraint rules were added to make the result hold up. Frames default to a single whole image; the README states you can ask for them to be split into separate frame module images by saying so explicitly in the prompt. During a run you see the intermediate images GPT produces, which is useful for spotting a bad segmentation before the deck is assembled.

Installing the skills into Codex and running a first deck

The README addresses installation to an AI agent rather than to a human, and the first step it gives is to send the GitHub repository URL to Codex and let it install the skills. The manual path is a copy into the Codex skills directory. Note that GordenSuperPPTSkill depends on the other two, so copy all three together, or copy the whole repository at once.

cp -R GordenImagePPTGen   "${CODEX_HOME:-$HOME/.codex}/skills/GordenImagePPTGen"
cp -R GordenImage2PPTX    "${CODEX_HOME:-$HOME/.codex}/skills/GordenImage2PPTX"
cp -R GordenSuperPPTSkill "${CODEX_HOME:-$HOME/.codex}/skills/GordenSuperPPTSkill"

The Python dependencies are three packages, installed with pip3.

pip3 install python-pptx pillow numpy

The image generation backend is resolved at runtime; the README says Codex uses its built-in imagegen, and points to references/runtime-notes.md inside each skill for details. For a first run, the README gives this prompt for image-only decks, with the model set to GPT 5.5 and reasoning effort on medium.

使用GordenImagePPTGen技能，生成一个N页的PPT，内容为XXX，要求PPT要求豪华、信息密度高、排版复杂

You should end up with one PNG per page plus an image-format .pptx. To convert an existing image instead, the README's prompt is:

把当前文件夹里的XXX.png，使用GordenImage2PPTX，还原成可编辑的PPT，必须严格遵循技能步骤

The instruction to follow the skill steps strictly is part of the prompt, which suggests the conversion degrades when the model improvises.

The quota cost is the real constraint, not the code

The README is unusually direct about the price of conversion: converting one image consumes roughly 10 percent of a Plus subscription's five-hour quota. That single sentence should drive most adoption decisions. A ten-page deck converted page by page is on the order of a full five-hour quota window, and the image generation pass is separate. This is not a tool you run over a 60-slide corporate template on a whim. The second limitation is host lock-in: the skills are written for Codex and depend on GPT image generation plus GPT vision, so there is no offline mode, no local model path, and no documented fallback. The third is licence ambiguity. The README grants commercial use with attribution, but the repository has no licence file, so the terms rest on a README sentence rather than a licence identifier. Where the README is silent matters too: nothing documents rollback, versioning, or how to recover a partially converted deck, and there are no releases to pin.

How this differs from python-pptx scripting and from template-based generators

The obvious alternative is writing python-pptx directly, which this project already depends on. The difference is direction of work. A python-pptx script starts from structured data and places shapes on a blank slide, so the visual ceiling is whatever layout code you write, and the output is editable by construction. GordenSuperPPTSkills starts from a generated image, which allows far denser and more elaborate layouts than hand-written layout code tends to produce, and pays for that with a vision-based reconstruction step that can misplace or merge elements. Template-based generators sit in between: they fill placeholders in a designed master, giving consistent editable output but constraining every slide to the template's slots. This project's trade is explicit: maximum visual freedom on the way in, uncertain fidelity on the way back. If your deck must survive a strict brand template, a template engine is the better tool. If you want a dense, designed-looking deck and can accept manual cleanup of the converted PPTX, the image-first route is the one this repository takes.

Maintenance, upgrade cost and what the licence actually says

The repository is not archived. The last push was on 2026-06-07, which is more than three months before today, so treat it as a project with no recent activity rather than one under active development. There are no releases, so there is nothing to pin and no changelog to read before upgrading. Upgrading means re-copying the skill directories, and because the skills are self-contained, a partial copy can leave GordenSuperPPTSkill calling a stale sibling. The practical upgrade cost is therefore low in effort but high in uncertainty: you are tracking the main branch. On licensing, the README states commercial use is permitted provided the GitHub source is credited or the author @Gorden Sun is marked. There is no LICENSE file in the repository listing, so the grant is a README statement, not a recognised licence identifier. That distinction matters if your organisation requires an SPDX identifier before adoption; this is a factual gap, not legal advice, and it is the first thing a reviewer should resolve.

Editorial conclusion

Adopt it if you already work inside Codex with a GPT 5.5 model, want image-first slide decks, and accept that converting one image costs roughly a tenth of a Plus subscription's five-hour quota. Do not adopt it if you need a CLI, a library, a self-hosted pipeline, or anything that runs without GPT image generation and vision. Before committing, verify three things: that the four-layer split (background, frame, icons and decoration, text) survives your own slide designs, that the quota cost per converted image matches what the README states, and that the licence terms are acceptable, because the repository states commercial use is allowed with attribution to the GitHub source or to @Gorden Sun but ships no licence file that a legal review can cite.

Frequently asked questions

Official sources

1. GordenSun/GordenSuperPPTSkills on GitHub
2. Issues
3. README