Garden Skills: a SKILL.md collection for Claude Code, Cursor and Codex

What Garden Skills actually ships

Garden Skills is a curated collection of Agent Skills, which the README describes as "production-ready" and targets at Claude Code, Cursor, Codex, and other AI coding agents. The unit of distribution is a skill folder under skills/, and each folder carries a SKILL.md file plus a manifest.json. The root package.json is explicit about this split: it is private, versioned 0.0.0, and its description says per-skill versions live in skills/<name>/manifest.json. The root package only ships maintainer scripts.

The four skills named in the README gallery are web-video-presentation, web-design-engineer, gpt-image-2, and beautiful-article. They are grouped by category: web video and presentation engineering, design and frontend, image generation and prompting, and "any source to beautiful article". That is the whole product. There is no runtime, no server, and no daemon. If you are looking for a library to import into an application, this is the wrong shape of project, because the deliverable is markdown instructions an agent reads, not code your program calls.

How a skill folder reaches your agent

The mechanism is file-based. A skill is a directory containing SKILL.md, and the agent reads that file to learn the workflow. Garden Skills does not define a new protocol; the README links to the Anthropic support article on what Skills are and to agentskills.io, so the loading behaviour belongs to whichever agent you use, not to this repository.

What the repository does own is packaging and release plumbing. The maintainer scripts under scripts/release/ are wired into npm scripts: cut-release.mjs cuts a release, pack-skill.mjs packs a skill (or all of them with --all), and update-readme.mjs syncs the README. The validate script chains list, pack:all and readme:check, which suggests the README's download links are generated rather than hand-edited. A skill is therefore a folder plus a manifest plus a zip artifact attached to a GitHub release. That is a clean design for a catalogue, but it also means the interesting behaviour lives one level down, in files the root README only summarises.

Installing Garden Skills with the skills CLI, a marketplace or a zip

The README lists five install routes: the skills CLI via npx, the Claude Code plugin marketplace, a pinned .zip from Releases, a manual copy into your project, and a git submodule. The README's table of contents names the first route as "skills CLI (npx)", and the repository's own package.json defines the maintainer scripts you can run once you have a checkout.

npm run list
npm run validate

Running list prints the skills the repository knows about; validate chains list, pack:all and readme:check, so it packs every skill and verifies the README is in sync. If either command fails, the checkout is not in the state the maintainer expects, and the release artifacts built from it would not be either.

The README also links a pinned release asset for web-video-presentation v1.2.2, published as web-video-presentation-1.2.2.zip. Download that asset from the Releases page and unpack it; you should end up with a folder containing SKILL.md and manifest.json.

Read SKILL.md before wiring anything up: it is the file the agent consumes, and it is the only place the per-step workflow is written down. The manual-copy and submodule options exist for people who want the files inside their own repository, which is the only route that gives you a normal git history for upgrades.

What web-design-engineer and web-video-presentation claim to do

These are the two skills with release entries, so they are the ones worth reading closely. web-design-engineer is aimed at landing pages, dashboards, prototypes, HTML slide decks, UI mockups and data visualisations. Its documented workflow is a sequence: understand product context, declare a design system, show an early v0, build the full experience, verify the result. Two details stand out. First, it converts the brief into a "five-dial Design Read" covering variance, motion, density, asset dependence, and brand fidelity. Second, it runs executable browser acceptance only when the user explicitly asks for acceptance, QA or browser testing, which is a deliberate restraint rather than an automatic step. It also ships a Design Direction Advisor with six schools and 25 style recipes, each with palette, typography, signature moves and anti-patterns.

web-video-presentation builds Vite + React + TypeScript presentations on a fixed 1920x1080 stage that scales to the viewport, driven by a (chapter, step) cursor and one narration beat per visual step. It documents 23 built-in themes and pluggable TTS with two built-in providers (MiniMax mmx-cli and OpenAI TTS via curl) plus a contract and snippets for ElevenLabs, edge-tts, Azure, Google Cloud and macOS say. Both skills pause at checkpoints before continuing. That checkpoint design is the most opinionated thing here: it trades speed for the ability to stop a wrong direction early.

Where the documentation runs out

The root README is a gallery. It tells you what each skill is for and links to a per-skill README, but it does not describe failure modes, rollback, or what happens when a skill's instructions conflict with your project's conventions. The README does not document rollback. It also does not state a minimum agent version, so compatibility is defined by whatever the linked agentskills.io and Anthropic pages say at the time you install.

Versioning is uneven. The release list shows web-design-engineer v1.3.0, web-video-presentation v1.2.2 and web-design-engineer v1.2.2, but nothing for gpt-image-2 or beautiful-article. Those two appear in the gallery without a matching release entry, so pinning them to a version may not be possible. The root package.json being private at 0.0.0 reinforces the point: there is no npm package version that tracks the collection as a whole.

There is also a naming problem you should plan around. Searches for this project collide with gardening. The related queries include "Gardening skills resume", "Basic gardening skills", "Gardening skill Sims 4" and "garden skills cheat sims 4", none of which have anything to do with an agent skill collection. If you are searching for help, add the agent name or "claude code" to the query.

Garden Skills versus copying prompts into your own repository

The realistic alternative is not another catalogue. It is keeping your own prompts and style guides in a folder your agent already reads. The difference is maintenance. With your own files you control every word and you know exactly when it changed. With Garden Skills you inherit 25 style recipes and 23 presentation themes that someone else maintains, which is a large amount of pre-written design judgement you would otherwise have to write yourself, at the cost of reading it before you trust it.

A narrower alternative is to take one skill and vendor it. Because the install options include a manual copy and a git submodule, you can pull in web-design-engineer alone and leave the rest. That avoids the uneven release coverage entirely and keeps the upgrade decision in your own commit history. The trade-off is that you stop receiving the maintainer's updates unless you pull them deliberately.

The repository was last pushed on 2026-07-12, and it is not archived. That is recent enough that the collection is not stale, but the release list shows the two design skills moving on their own cadence, so treat each skill as a separate dependency rather than one product.

Licence and upgrade cost

The repository is MIT licensed, and the root package.json repeats "license": "MIT". For most teams that means you can copy, modify and redistribute the skill files, including inside a commercial repository, provided you keep the licence notice. This is not legal advice; if you redistribute the skills as part of a product, have your own counsel read the LICENSE file rather than this article.

The upgrade cost is where MIT is least helpful. Because the root package is private and versioned 0.0.0, there is no single version number to pin. Upgrades happen per skill, through skills/<name>/manifest.json and the matching release tag, and the two skills without release entries have no tag to pin to. The maintainer scripts give you a way to check consistency (npm run validate runs list, pack:all and readme:check), but that validates the repository, not your installation. If you vendor a skill, record the release tag in your own commit message; that is the only rollback record you will have, because the README does not describe one.

Editorial conclusion

Adopt Garden Skills if you already run a SKILL.md-compatible agent and want a starting point for design, presentation or image work rather than a blank prompt. Do not adopt it if you need a stable API surface, a versioned release channel for every skill, or documented rollback: only web-video-presentation and web-design-engineer appear in the release list, and the top-level package.json is private at version 0.0.0. Before installing, open skills/<name>/SKILL.md for the skill you actually want and read the manifest, because the root README describes the gallery more thoroughly than it describes any single skill's behaviour.

Frequently asked questions

Official sources

1. Official documentation
2. Official README
3. Project repository
4. Release notes