Kangarooking Skills ships sixteen workflows as plain directories

Sixteen directories, each one a SKILL.md

The repository is a container, not a library. Its stated purpose is to publish the agent skills the author has actually got working in his own workflows, and its unit of distribution is a directory.

Each directory is a structured workflow an agent can read directly, and every one contains a SKILL.md file. The more complex tasks ship more than that: scripts to run, reference material to consult, and templates to reuse. That is the whole packaging model, and it is why the repository needs no build step and no runtime of its own.

The catalogue is grouped by what the skills are for. Image generation holds a single-call generator and a multi-agent pipeline. Visual design holds the cover workflow and a paper-zine look. Brand design derives a system from a real shop. There is 3D generation, a motion website builder, a content production pipeline, two content growth skills, an information monitor, an asset processor, a book illustration workflow, two agent engineering harnesses, and a personal growth skill.

The project targets Claude Code, Codex, and any other agent that supports the Agent Skills open standard. The repository publishes no releases, and the last push to main is dated September 7, 2026.

Install by asking the agent, or copy the directory

There are two installation routes, and the first one is a single sentence to the agent.

In a tool that supports agent skills, you ask it to install the skill by giving the URL of that directory in the repository, and the agent does the rest. You name the directory, so installing one skill means installing that folder rather than the whole repository.

The manual route is a clone and a copy, with a different destination folder per tool:

git clone https://github.com/kangarooking/kangarooking-skills.git
mkdir -p ~/.codex/skills
cp -R kangarooking-skills/<skill-name> ~/.codex/skills/
mkdir -p ~/.claude/skills
cp -R kangarooking-skills/<skill-name> ~/.claude/skills/

Both destinations are a hidden skills folder in the home directory, one for Codex and one for Claude Code, and you copy a named subdirectory rather than the repository root.

There is a third path for tools that cannot install skills at all: put the directory's SKILL.md and whatever resources it references into your project, and let the agent follow the process described inside. Since the deliverable is a document an agent reads, that fallback works, and it is the reason there are no scripts that must be compiled before the instructions mean anything.

Image skills: one call versus a multi-agent pipeline

The two image generation skills sit at opposite ends of the control spectrum, and the difference shows in what each one asks of you.

The direct generator is a single call to a hosted image model through an API service. It handles text to image, reference images, resolution and ratio control, and it waits for the result and downloads it without you polling. It supports several resolution tiers and the common aspect ratios, and it accepts URLs, local images, and several reference images at once. Its one explicit engineering decision is about credentials: the API key is read from environment variables only, specifically so it is never written into code or into a commit.

The multi-agent version is a workflow rather than a call. It organises design analysis, reference image selection, generation, and series consistency into stages, and it carries a built-in design compilation capability and a case library it can reuse. It is built for batch generation and for producing a set of images that look like one set, and it is described as suited to agent environments that expect several rounds of interaction.

The practical difference is iteration. The first is a request and a download. The second is a pipeline you are inside.

cover-skill runs two stages with fixed platform ratios

The cover skill is the most specified workflow in the repository, and the specification is the point.

It is built around a creator's real portrait, a kangaroo mascot, a logo and product assets. Stage one generates a key visual of the person and the kangaroo interacting naturally, then outputs four directions, labelled A through D, natively at whichever platform or ratio you specified. The ratios are not left to taste: 21:9 for WeChat articles, 16:9 for Bilibili, 9:16 for Douyin, and 16:9 only as the default when you name neither. Every direction has to preserve the real person's identity, the kangaroo character, and a centred subject, and directly cropping or pasting the original mascot image is forbidden.

Stage two takes the direction you chose and adapts it natively to five formats, adding 4:3 landscape and 3:4 portrait to the three platform ratios.

Two implementation details make the output usable rather than merely pretty. The layout is deterministic, which is what preserves Chinese copy and a real logo instead of letting a generative pass redraw them. And there are per-image QA and export validation scripts, so the check happens after generation rather than in your eyes.

The suggested phrasing shows the intent: produce four cover directions first, let a person choose, then adapt across platforms.

Two skills that start from a photograph

The paper-zine skill and the brand system skill both take photographs as their input, and both are explicit about what must survive the transformation.

The paper-zine skill turns ordinary photos into quiet, restrained editorial pages. It recomposes with torn paper, generous whitespace, simplified line art, and a single structural colour, while preserving the relationships inside the original scene. It covers everyday scenes, people, pets, food, interiors, streets, architecture and travel. The key rule is a distinction between pixel fidelity and generative recomposition: anything identity-critical defaults to fidelity rather than being redrawn. Delivery comes as a single poster, a three-image series, a series cover, or a side-by-side comparison with the original.

The brand skill starts from real shop photographs and does something more commercial. It diagnoses the operating and spatial relationships first, then extends explainable visual concepts into a brand system covering the storefront, packaging and social material. It keeps real spatial anchors such as doors, windows, the entrance, displays and building materials, derives a logo and supporting graphics from the shop structure, product silhouettes and the purchase action, and selects materials along the customer journey rather than a generic list. The logo derivation and before-and-after evidence are kept.

Both skills are arguing the same thing: the photograph is a constraint, not a starting sketch.

3D generation and motion sites pick defaults deliberately

The 3D skill generates downloadable models from text, from an image, or from several views, using a Hunyuan-based service. It supports physically based materials as well as white models, sketch-based generation and automatic topology, and the example request is specific: a GLB model with PBR materials from a picture of a dragon.

Its default transport is worth noting. Rather than calling the vendor's own client, it goes through a TokenHub endpoint that speaks an OpenAI-compatible interface, and only falls back to the vendor's own SDK if you ask it to explicitly. That is a portability decision: the skill is written against an interface shape rather than one company's client library.

The motion website builder does something different with the same kind of intent. It turns product material into a site driven by scrolling, with a cinematic feel, delivered so you can preview it locally. It generates a continuous visual story using two named generation systems, then does the unglamorous work: checking the real first and last frames, handling transition segments and short dissolves, and checking continuity between clips. The output is a Vite and React site with a mobile fallback and a local preview package.

It also credits the project it learned the scroll video and boundary frame techniques from, and keeps a full MIT attribution notice.

The growth skills keep their evidence

Three skills are about finding and writing for an audience, and all three are designed to leave a record rather than a vibe.

The topic skill sweeps several content platforms for recent breakouts and for the awkward case of a low-follower account with strong performance, which is the sample that tells you something transferable. It covers WeChat articles, X, Bilibili and YouTube, calls the collection logic appropriate to each platform, and keeps the source link, the account metrics, the time window and the evidence for each hit. The value is in those fields: a topic you cannot trace back to a post is not a finding.

The title skill generates candidates in bulk and then does the harder part, recommending which one is worth publishing. It covers the same four platforms, supports searching a local title library and reusing a similar structure, and records the choice and the feedback so the method iterates instead of resetting.

The monitor skill fetches updates for specific accounts and outputs structured data. It accepts a user ID, a handle or a profile URL, returns JSON or CSV with engagement metrics, and can sync into a multidimensional table, with guidance for running it on a schedule.

Together they form a loop: find, name, watch.

Asset workflows and the two harness skills

The remaining skills split into pipelines for existing material and pipelines for managing an agent's own work.

The video skill turns a platform link into a package: the original video, the platform's own caption, an audio transcription, and standardised metadata. It covers several Chinese and global platforms, and it distinguishes the platform's original caption from a speech-to-text transcription of the audio, which are different documents with different uses. One platform currently requires an extra step through a small program before the download works.

The book illustration workflow strings together the tedious parts of putting a chapter together: a master table of screenshots and the prompts to produce them, consistent figure numbers and filenames, insertion points in the text, removal of author notes, and a sync.

The two agent engineering skills are about control. One initialises a plan, build and verify framework for a project, supplying agents, commands, hooks and documentation structure for work that needs self-verification over a long cycle. The other breaks a complex requirement into a task system with acceptance criteria, a single state file, emphasis on tests and commits, and resumable execution, aimed at multi-session work where an agent that loses the thread is the failure mode.

A related repository takes a book and distils its method into triggerable, composable, testable skills, which is the upstream half of this collection.

Editorial conclusion

Kangarooking Skills fits someone who prefers a workflow they can read and edit over a tool they configure, and who already uses an agent that supports the Agent Skills standard. It does not fit anyone who needs a maintained library with a release cadence, since the repository publishes no releases and declares no licence. Before you copy anything, read the SKILL.md rather than the summary line in the table, because several of these skills encode a specific platform ratio or a specific fidelity rule, and a skill that quietly picks a default you did not want is worse than no skill at all.

Frequently asked questions

Official sources

1. Issues
2. kangarooking/kangarooking-skills on GitHub
3. README