ClaudeKit Skills: a catalog of agent workflows, read closely

What an Agent Skill is, according to the README's own framing

The README opens by defining the format rather than the product. Agent Skills are specialized workflows that empower Claude to perform complex, multi-step tasks with precision and reliability, and they combine mission briefs, guardrails and integration hints.

That three-part definition is worth holding on to, because it explains what you are actually installing. A skill is not a library and not a prompt template. It is a bundle of three things with three different jobs: a mission brief that states what outcome is wanted, guardrails that constrain how the work is done, and integration hints that point at the specific tools, libraries and conventions in play. The stated purpose is to turn generic AI assistance into disciplined automation.

The README links to the official Agent Skills overview documentation for the format itself, which tells you this catalog is a consumer of a documented convention rather than a proprietary invention. That distinction matters when you are evaluating it, because you can read the spec, decide the format is right for your setup, and then judge each individual skill on its own content.

The project describes itself with a single sentence, all powerful skills of ClaudeKit.cc, and points at a home page on a cc domain. The language recorded for the repository is Python, though the tree suggests most of the payload is documentation rather than runnable code, which is the normal shape for a skills catalog.

The catalog, grouped by the kind of work it covers

Eleven skills are named in the README, and they are grouped under six headings. Reading the groupings tells you what the author considers separable disciplines.

Under authentication and security there is better-auth, described as a TypeScript authentication framework supporting email and password, OAuth, 2FA, passkeys and multi-tenancy, and noted as working with any framework including Next.js, Nuxt and SvelteKit.

Under AI and agent development there is google-adk-python, covering Google's Agent Development Kit for tool integration, multi-agent orchestration, sequential, parallel and loop workflow patterns, and deployment to Vertex AI or custom infrastructure. Next to it sits context-engineering, which is the most substantial entry in the catalog by description. It covers context fundamentals, degradation patterns, optimization through compaction, masking and caching, compression strategies, memory architectures, multi-agent patterns, LLM-as-Judge evaluation and tool design, and it tells you to use it when designing agent architectures, debugging context failures or optimising token usage.

Under backend development, backend-development spans Node.js, Python, Go and Rust with NestJS, FastAPI and Django, PostgreSQL, MongoDB and Redis, REST, GraphQL and gRPC, OAuth 2.1 and JWT, OWASP Top 10, caching, sharding, Docker, Kubernetes and CI/CD. That is a very wide surface for one skill, and it is the entry most likely to overlap with whatever standards your team already has.

Under AI and machine learning, ai-multimodal processes and generates multimedia through the Google Gemini API. It analyses audio including transcription and summarisation up to 9.5 hours, understands images through captioning, object detection, OCR and visual question answering, processes video through scene detection and question answering with YouTube URLs up to six hours, extracts from documents including PDF tables, forms, charts and diagrams, and generates images from text or by editing and composition. It notes Gemini 2.5 and 2.0 support with context windows up to two million tokens. Context-engineering appears a second time here, in a shorter form.

Under design and aesthetics there is aesthetic, which is organised around four named stages, BEAUTIFUL for aesthetic principles, RIGHT for functionality and accessibility, SATISFYING for micro-interactions and PEAK for storytelling. It also describes a workflow of capturing inspiration screenshots with chrome-devtools and ai-multimodal, then generating design images iteratively until the standard is met.

Under web development there are five entries. web-frameworks covers Next.js App Router, Server Components, RSC, PPR, SSR, SSG and ISR, plus Turborepo for monorepos and RemixIcon for icons. ui-styling covers shadcn/ui on Radix UI and Tailwind, plus canvas-based visual designs, themes and dark mode. frontend-design is a shorter entry about producing distinctive production-grade interfaces that avoid generic AI aesthetics. frontend-development covers React and TypeScript patterns including Suspense, lazy loading, useSuspenseQuery, a features directory for file organisation, MUI v7, TanStack Router and TypeScript best practices. And threejs covers Three.js work with WebGL and WebGPU, custom shaders, PBR materials, VR and XR, games and data visualisation.

One detail is easy to miss and worth flagging. The README ends mid-sentence in the web development list, and context-engineering is catalogued under two different headings with two different lengths of description. Both suggest the README and the catalog behind it drifted apart as skills were added.

The plugin marketplace path, and why it is the documented one

The README has a December 2025 announcement that changes how you install this. ClaudeKit Skills is available as a Plugin Marketplace, with installation and automatic updates, and the commands are:

/plugin marketplace add mrgoonie/claudekit-skills
/plugin install ai-ml-tools@claudekit-skills

Two things about that snippet are instructive. The first line adds the repository as a marketplace source, and the second installs a named plugin from it using an at-sign reference. The example package is called ai-ml-tools, which tells you the repository publishes multiple plugins rather than one, and that the plugin names do not match the skill names one to one. You will need to see the marketplace listing to know which plugin carries which skills.

The announcement frames this as replacing manual git clones, which is the correct trade. A marketplace install gives you update handling without remembering to re-pull, and it means the skill content is versioned by the plugin rather than by whatever your working copy happens to contain.

The README also points to a Getting Started section further down, and the repository has a separate `INSTALLATION.md` at the root, so there is a manual path documented alongside the marketplace one. That pairing is useful: if your environment cannot reach a marketplace, the file walkthrough is the fallback, and if you want to inspect a skill before installing it, you can clone and read.

One warning about the marketplace approach follows from how these skills work. Installing a plugin means its instruction text is loaded into your assistant's context for sessions where it applies. That is the mechanism by which a skill helps, and it is also the reason a catalog this broad is worth installing selectively rather than wholesale.

What the tree contains, and what each directory is for

The repository root has eleven entries, and the names fall into groups that are easy to separate.

There are `.claude/` and `.claude-plugin/`. The first is where the skills themselves live, since every catalog entry links to a path like `.claude/skills/<name>`, so that directory is the payload. The second is the plugin manifest directory, which is what makes the marketplace install work. There is also `plugins/`, which likely holds the individual plugin packages that the marketplace publishes, consistent with the `ai-ml-tools` example in the install command.

Then there is the documentation layer: `INSTALLATION.md`, `MCP_MANAGEMENT.md` and `REFACTOR.md` at the root, plus a `docs/` directory. Each of those three root files is linked from the README as the destination for one of its three dated announcements, which is a tidy structure. Each announcement is short and links out to exactly one file, so the README stays a catalogue and the reasoning lives elsewhere.

There is `CLAUDE.md` at the root, which in this ecosystem is a project-level instruction file. Its presence suggests the repository dogfoods the format it publishes, though the file itself is not quoted in the README so its contents are not documented here.

`plans/` and `assets/` round out the tree. The README references an asset image for the November 2025 announcement, so `assets/` holds screenshots. `plans/` is not referenced anywhere in the README, which makes its purpose ambiguous from the documentation alone; it may hold planning documents for skills under development.

There is a `claudekit.png` logo, a `.gitignore`, and no releases. The absence of a published release means there is no version to pin, so a marketplace install tracks whatever the main branch contains.

Three dated announcements that describe three refactors

The most informative content in this README is the sequence of dated notes near the top, because each one names a specific architectural change rather than a feature.

The November 2025 note is about context bloat. Its title names two things it removed: an mcp-manager subagent and mcp-management skills, and the accompanying image and `MCP_MANAGEMENT.md` link invite you to read more. The framing is that the previous design consumed context unnecessarily, which is the exact failure mode that context-engineering, one of the catalog's own skills, is written to solve. It is a little ironic that the fix involved deleting a skill from the catalog, and worth keeping in mind as a reminder that a skill collection accumulates rather than curates unless someone actively prunes it.

The refactor note is separate and vaguer. It announces that the skills have been refactored and points at `REFACTOR.md`. Without that file, what changed is unclear from the README alone, but the existence of a dedicated document suggests a structural change rather than a rewrite of individual skills.

The December 2025 note is the plugin marketplace, discussed above.

Three refactors in roughly two months, each with a changelog entry, is the signature of an actively reworked project rather than a static list. It also means the shape of the skills has changed more than once, which is a reason to read the current `REFACTOR.md` before assuming a skill has always looked the way it does now.

A fourth detail sits below the notes and above the catalog: a link to a Claude Code marketplace reference. Combined with the Dec 2025 announcement, it confirms the project has moved from being a collection you clone to being a marketplace you subscribe to.

Ecosystem context and a maintenance caveat worth stating plainly

The README lists six other projects by the same author, and they are worth reading as a set because they show what this author's projects are for.

There is a blog on a personal domain, a Claude Code skills library at uupm.cc which is the UI UX Pro Max skill, a knowledge base described as being for both humans and AI agents, an OpenClaw project for enterprise, a shipping solution, and the ClaudeKit storefront itself. The pattern is one author building a set of small, focused tools around agent workflows, each on its own domain, with skills as one of the output formats.

That context explains the shape of this repository. It is a distribution point for one category of that output, not a framework with an API. If you are evaluating whether to adopt ClaudeKit Skills, you are evaluating a catalogue and its install path, not a dependency you will call from code.

Now the caveat. The main branch of this repository was last pushed on 2026-04-03. That is more than six months before the present date. There are no published releases, so nothing has been cut since then either. The December and November announcements in the README are dated 2025, which means the most recent documented changes predate that push by several months.

That does not make the skills bad. Documentation written in late 2025 about Agent Skills, Next.js App Router, shadcn/ui on Radix, MUI v7, Three.js with WebGPU and Google ADK is broadly current. But it does mean the repository is not actively receiving fixes, and if you install from the marketplace expecting automatic updates, the updates are updates to a branch that has been quiet for a while. Read a skill before you depend on it, and prefer copying a skill into your own project over referencing the catalogue, so that what you rely on is under your own version control.

Editorial conclusion

ClaudeKit Skills is worth adopting selectively rather than wholesale, because the catalog covers eleven named areas with real depth in each, and installing all of them puts a lot of instruction text into every session you start. The plugin marketplace path is the right one to use rather than a manual clone, since it handles updates, and the skills directory under `.claude/` is where you can read what a skill actually contains before you commit to it. Check the push date before you rely on it, because the main branch here was last updated on 2026-04-03, well over six months before this article was written. If you need a maintained dependency for a production pipeline, treat this as reference material to read and adapt rather than something to pin.

Frequently asked questions

Official sources

1. Issues
2. mrgoonie/claudekit-skills on GitHub
3. Project website
4. README