xiaobei-skill: an index root with no SKILL.md, three skills, and a requirements file that lists two of them

The root has no SKILL.md on purpose and cannot be installed as one skill

The most important line in the page is an instruction not to install anything yet. This repository is a collection index, and the root deliberately contains no SKILL.md, so the root cannot be installed as a single skill. If a user asks to install `xiaobei-skill` without naming a child, the page instructs the host to read `skill-catalog.json`, show the list, let the user choose, and install only the selected subdirectory, with an explicit rule not to guess and not to install everything at once. Each subdirectory under skills/ is therefore the unit of installation, and the page states that each one is self-contained and does not need files from the repository root in order to work. Manual installation is a clone plus a symlink into the host's skills directory, swapping the last directory name for whichever skill you want.

git clone https://github.com/xiao24bei/xiaobei-skill.git
mkdir -p ~/.codex/skills
ln -s "$(pwd)/xiaobei-skill/skills/xiaobei-skill-image-to-vba" ~/.codex/skills/xiaobei-skill-image-to-vba

One caveat is written into the same paragraph: `skill-catalog.json` is this repository's own machine-readable index and is not the official OpenAI Skill schema, so anything reading it should not assume it conforms to a vendor specification.

The root requirements.txt lists two of the three skills

The repository ships a top-level requirements.txt, and it contains two include directives rather than three:

-r skills/xiaobei-skill-image-to-vba/requirements.txt
-r skills/xiaobei-skill-academic-paper-to-ppt/requirements.txt

The third skill, the PowerPoint reconstruction one, has no line there, even though its own directory holds a requirements.txt according to the structure listing. That leaves two documented paths that disagree. The dependency section instructs you to install a specific skill's script dependencies one at a time with pip, listing all three directories explicitly, while the root file quietly omits one. Anyone who installs from the root file and then runs the PowerPoint skill gets a missing dependency error rather than a clean import failure, and the fix is to install that directory's own requirements.txt by hand.

Two naming systems, and the demo uses a third name that is not published

Invocation names and display names are separate on purpose. The internal call names use lowercase letters and hyphens for install and invocation compatibility, giving $xiaobei-skill-image-to-vba, $xiaobei-skill-rebuild-image-in-powerpoint and $xiaobei-skill-academic-paper-to-ppt, while the display_name shown in the Codex interface carries a personal brand prefix on every entry. That is a defensible choice for a curated collection, but it means a user has to learn two strings per skill. The naming has already shifted once: the demonstration video was recorded while the PowerPoint skill was still called `$academic-image-to-ppt`, and the page says the public name is now the $xiaobei-skill form and that installation and invocation should follow the public name. A viewer who types the name they saw in the recording gets a name that was never published.

A 56 second recording claims about 240 editable objects

The single GitHub release in this repository is a demonstration asset rather than a version of the software: rebuild-image-in-powerpoint-demo-20260827, published on 2026-08-27, carrying a 56 second MP4 that shows the reconstruction workflow. The page describes what the recording contains, with PowerPoint kept visible on one side drawing region by region and object by object while the host works on the other, starting from a blank slide and ending with a complex biological mechanism diagram that keeps about 240 editable objects. Two details are worth repeating because they are engineering decisions rather than marketing. The video is deliberately kept out of the skill directory and out of the repository history so that installing one skill does not pull roughly 52 MB of demo material, and the page states that the reference image and the paper content belong to their original authors and rights holders.

The VBA path makes the Office trust setting part of your setup

The Image to VBA skill converts PNG and JPG content into editable Office shapes and VBA code rather than repainting the picture as a static image, which means the output is executable. The page handles this directly: run only VBA you trust, read the generated .bas file first, and expect Office to block macro import or to ask you to enable trust access to the VBA project object model, which it describes as a normal security boundary rather than a defect. The same section adds a WPS caveat, since the skill does not assume WPS will import or run macros automatically and gives a manual fallback instead. Combined with the rebuild skill, which drives PowerPoint itself on Windows, the picture is a workflow where a lot of trust sits with the host application, and the documentation at least names every place where the host may refuse.

Hybrid mode crops the input image and records it in the report

The reconstruction approach has a documented cost, stated in both a technical and a legal section. Text, arrows, boxes, tables, legends and axes are rebuilt as native Office shapes where possible, but micrographs, photographs, complex 3D renders and logos are kept as small raster crops, and the report marks where that happened. The pipeline is manifest-driven, listing the elements first and building the coordinate model afterwards to reduce missing elements and misaligned arrows, and it closes the loop when a local Office install is available by exporting a render and diffing it against the source image. The rights note follows directly from the crop behaviour: the output embeds pieces of the input, so the person running the skill needs the right to use the source image, and the page recommends self-drawn, public domain or licensed material for examples.

The structure listing stops before the third skill

The tree at the end of the page is the most convenient summary of the layout and it is also incomplete in two visible ways. It lists the image-to-vba directory with SKILL.md, an agents/openai.yaml, requirements.txt, assets/, references/ and scripts/, then lists the PowerPoint reconstruction directory with the same set minus assets/, and then the third entry begins and stops after the letters SKILL.m. So the listing shows one directory with an assets folder, one without it, and one that is cut off, which means the tree cannot be used to compare the three skills. The shape of each directory is still clear from the first entry: a SKILL.md as the unit of installation, an agents/openai.yaml, a requirements.txt for script dependencies, and references and scripts directories for the material the skill loads.

Part of the index page is audience recruitment rather than documentation

Before any technical content, a large share of the page is spent on the author's audience. It opens with an introduction from a second-year master's student in electronic information, states a follower count in the tens of thousands on a short video platform, and invites readers to add a WeChat contact, with a suggested convention for what to write in the friend request so that requests can be sorted by school, field and year. It also links a personal academic site and a research API relay service. None of that is wrong for a public repository, and the repository does hold real assets alongside it, with AGENTS.md, CONTRIBUTING.md, SECURITY.md, NOTICE, CITATION.cff and a validation script. It does mean the page cannot be read as documentation end to end, and a reader looking for how a skill behaves has to step into the subdirectory files.

Editorial conclusion

This is a prompt-and-script collection aimed at a specific host, the Codex skill runtime, and it works best when you already know which of the three skills you want, because the index root is deliberately not installable and the page asks the host to read a catalog file and ask you first. Two things deserve a look before you adopt it. The top-level requirements.txt pulls in only two of the three skills, so a plain install of the collection silently misses the PowerPoint reconstruction skill. And the Image to VBA path produces VBA macros, which means the Office trust settings become part of your setup, on top of a documented WPS fallback. Read the per-skill SKILL.md files rather than the index page, and treat the demo claims as demonstration: about 240 editable objects in a 56 second recording is a recorded example, not a benchmark.

Frequently asked questions

Official sources

1. License: Apache-2.0
2. Project website
3. README
4. Releases
5. xiao24bei/xiaobei-skill on GitHub