GordenPPTSkill: a python-pptx skill that fills 17 Chinese PPTX templates instead of drawing slides

The problem GordenPPTSkill targets: layout survives, text changes

Ask a language model to produce a slide deck and you usually get one of two things. Either it emits a Markdown outline that a renderer turns into plain boxes, or it writes python-pptx code that positions every shape by coordinate. The first looks generic. The second breaks the moment the text is longer than the author expected, because nothing in the generated code knows how much room the original designer left.

GordenPPTSkill takes a third route. The layout is fixed in advance, by a human, inside a .pptx file. The agent's job is only to decide what text goes into which placeholder. The README describes the result as decks with high information density and complex layout, aimed at state-owned enterprises and large internet companies, and says the same skill also produces simple business-style decks. That is a claim about the bundled templates, not about a layout engine.

The intended user is someone who already drives an AI agent and wants a PPT skill installed into it. The README points that reader at SKILL.md, the AI entry document, and separately at templates/INDEX.md for template categories and recommendations. The repository is personal and research oriented: the README states in bold that the repository and its bundled templates are for personal study and research only, and that commercial use is forbidden.

How the build pipeline works: edits.json in, a real .pptx out

The repository layout tells most of the story. scripts/ holds five user-facing scripts covering build, render, update and manifest work. templates/ holds 17 templates, each made of four files. references/ holds editing rules, a schema and workflow notes. SKILL.md is the entry document an agent reads, and manifest.json records a version and sha256 per file.

The build script takes four arguments in the README's example: a template .pptx, an edits.json, an output path, and a --detail flag pointing at that template's detail.json. So the data flow is: the agent reads SKILL.md and the template's detail.json to learn which shapes exist and what each one is for, writes edits.json describing the text for those shapes, then build_pptx.py opens the template with python-pptx and writes the text in. The README calls this non-destructive text-only editing, and the description of the project repeats the phrase: text-only, layout intact.

Two optional dependencies sit outside that core. Rendering a preview needs LibreOffice (soffice) and poppler (pdftoppm). Neither is required to produce the .pptx. The update mechanism is the other moving part: VERSION, CHANGELOG.md, updates.json and manifest.json together let the skill update itself, and the README describes this as the skill updating like software when new selectable templates are published.

Installing GordenPPTSkill and building a first deck

There is no package to install. The README's usage section says to send the repository URL to your agent and ask it to install the skill, then ask for a deck. For anyone working from a shell, the README gives an explicit quick start.

First confirm the dependencies. python-pptx must be 1.0 or later, and the two rendering tools only matter if you want preview images.

# 1. 确认依赖
python3 -c "import pptx; print(pptx.__version__)"   # python-pptx 1.0+
soffice --version    # LibreOffice (仅渲染预览时需要)
which pdftoppm       # poppler   (仅渲染预览时需要)

Then pick a template, write edits.json for it, and run the build. The README's example uses the minimal-business-summary template and passes its detail.json so the script knows the shape contract.

# 2. 选定模板 + 写 edits.json，跑构建
python3 scripts/build_pptx.py \
    templates/minimal-business-summary/template.pptx \
    edits.json \
    out/final.pptx \
    --detail templates/minimal-business-summary/detail.json

What you should see is out/final.pptx, a real PowerPoint file with the template's layout intact and your text in the shapes. If you also want to check the result visually, render it:

# 3. (可选) 渲染最终预览图
python3 scripts/render_slides.py out/final.pptx out/preview --dpi 144

The README does not document the edits.json schema itself; it points to references/ for editing rules and schema, so that is where to look before writing your first file. One environment step is easy to miss. The templates use 微软雅黑 heavily, and if the machine lacks it the README suggests a fontconfig alias mapping it to WenQuanYi Micro Hei, DengXian, Noto Sans SC or PingFang SC, or installing Noto Sans SC with brew install --cask font-noto-sans-sc.

Where GordenPPTSkill stops being the right tool

The licence is the first hard boundary. The README states the repository and its bundled templates are for personal study and research only, commercial use is forbidden, and redistribution is forbidden. The repository's LICENSE file carries a NOASSERTION identifier, which means the licence could not be classified automatically. Anyone planning to put a deck in front of a client should read LICENSE and NOTICE.md directly rather than relying on the README's one-line summary.

The second boundary is the template count. There are 17, and the agent cannot invent an eighteenth. If your content does not match the shape of a bundled template, you are editing around the mismatch, not solving it. The README acknowledges this indirectly by offering custom private templates through a WeChat contact, which is a different arrangement from the open repository.

The third is the text-only constraint. Because the build writes text into existing shapes and does not reposition them, a paragraph that runs long has nowhere to go. The repository does not document automatic overflow handling, and the README does not document rollback if a build produces a bad deck; you re-run the script against the template again. Finally, this is a Chinese-first project. The templates, the font assumptions and the documentation all assume a Chinese-language deck and a machine with Chinese fonts configured.

GordenPPTSkill compared with code-generating slide tools

The closest alternative in kind is a tool that has the model write python-pptx or similar layout code directly, generating shapes and coordinates rather than filling a fixed file. The difference in approach is where the design decisions live. In a code-generating tool, the model decides both content and geometry, so you can get a layout no template anticipated, and you can also get overlapping boxes and inconsistent margins on every run. In GordenPPTSkill, geometry is frozen in the .pptx and the model only supplies strings, so output is consistent across runs but bounded by what the templates contain.

That trade favours GordenPPTSkill when the deck is a recurring format, a quarterly review or a project introduction, and you want the same visual grammar each time. It favours the code-generating approach when the content shape is genuinely new and no bundled template is close. The README's own framing supports the recurring-format reading: it advertises templates in several styles and says the skill suits state-owned enterprises and large internet companies, which are environments where decks follow house conventions.

A second difference is model compatibility. The README states the skill works with DeepSeek, Xiaomi Mimo, Claude and GPT, and that domestic Chinese models also perform well. Because the agent only produces edits.json, the model never needs to reason about coordinates, which is a plausible reason the README can make that claim across such a range of models.

Maintenance, versioning and what an upgrade actually costs

The last push to the default branch was on 2026-06-22. The repository is not archived. There are no retrieved releases, so version tracking runs through VERSION (1.0.0 in the README's directory listing), CHANGELOG.md for humans and updates.json for machines.

The upgrade design is the interesting part. manifest.json records a version and sha256 for each file, and the README describes an automatic update mechanism: when new selectable templates are published, using the skill updates it. That means the set of available templates can grow under you. The cost is that a template you built against can change, and the README does not document how a local edit to a template survives an update, nor how to pin to a specific template version. The sha256 entries in manifest.json are the mechanism you would use to detect that a template changed, but the README does not spell out a verification workflow.

On licence cost: the README's non-commercial restriction propagates to anything you build with the bundled templates, since the repository states it holds no copyright over the PPT templates and forbids redistribution. Open source dependencies are LibreOffice, python-pptx, Poppler and WenQuanYi Micro Hei. This is a description of what the repository states, not legal advice.

Editorial conclusion

GordenPPTSkill fits people who want a finished, layout-complex Chinese deck from an AI agent and are willing to work inside one of the 17 bundled templates, on a machine that has python-pptx 1.0+ and the 微软雅黑 font available. It does not fit anyone who needs a commercial deliverable, a template of their own, or a layout the agent invents from a blank canvas. Before committing, run the dependency check from the README, open templates/INDEX.md to confirm a template matches your content shape, and read the repository's own licence files rather than the README's summary, because the LICENSE file carries a NOASSERTION identifier and the README states the repository holds no copyright over the PPT templates.

Frequently asked questions

Official sources

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