Most "skills collections" on GitHub are indexes. ConardLi/garden-skills is the other shape: five fleshed-out Agent Skills with galleries, pinned .zip releases, and install paths for Claude Code, Cursor, Codex, Gemini CLI, and OpenCode.
The README calls it "a curated collection of production-ready Agent Skills." GitHub listed 11,345 stars and 1,412 forks on August 28, 2026. Created April 21, 2026; last code push July 12, 2026. That is enough traction to change what you install this week — if you pick the right skill and ignore the stale "install all four" line.
If you still need the mental model, start with what agent skills are. This post is the practical follow-up: install, choose, compare to rolling your own and to the explainx.ai skills registry.
TL;DR — questions people type after the GitHub page
| Question | Direct answer |
|---|---|
| What is it? | Five MIT SKILL.md packs from ConardLi — design, 16:9 video decks, GPT Image 2, local KB retrieval, editorial articles. |
| Who is it for? | Builders who already run a coding agent and want opinionated visual / retrieval workflows, not a general code-review library. |
| Install? | npx skills add ConardLi/garden-skills (or -s <skill>). Claude Code also has /plugin marketplace add ConardLi/garden-skills. |
| How many skills? | Five in skills/. README still says "install all four" in one snippet. Plugin packs cover four of five. |
| API keys? | Not for install. gpt-image-2 and optional TTS need provider credentials when you generate. |
vs explainx.ai /skills? | Garden is one author's deep pack. /skills is the ranked corpus across authors. Use both. |
| License? | MIT © ConardLi. |
What is garden-skills, and who is it actually for?
Garden Skills is not a marketplace and not a thousand-skill awesome-list. It is a small garden: each skill is a folder with SKILL.md, a human README, optional references/, scripts/, and assets/ — the layout the Agent Skills spec expects.
Who should clone it:
- You generate landing pages, dashboards, or prototypes and keep getting generic AI UI.
- You need recordable 16:9 web presentations (Vite + React), not a slide SaaS.
- You already call GPT Image 2 (or a compatible image API) and want structured templates plus three runtime modes.
- You have a local
knowledge/tree (Markdown, PDF, Excel) and want retrieval that does not dump whole files into context. - You want source → share-ready article (HTML, optional PDF) with hard editorial checkpoints.
Who should skip it:
- You need CI, code review, migrations, or deploy playbooks — this pack is design / media / retrieval. Browse
/skillsor write a thin SKILL.md for those. - You wanted a one-click "make my whole agent smart." Five overlapping visual skills will compete on triggers if you install them all.
The web-design-engineer frontmatter is explicit about not stealing sibling jobs: it is for visual front-end artifacts, "not for back-end, CLI, non-visual coding, source-to-longform article conversion, or narration-driven click-through video presentations." That routing is the garden's real product — five skills that know when to stay out of each other's way.
How do I install garden-skills into Claude Code, Cursor, or Codex?
Five install paths are documented. Use A unless you have a pinning or air-gap reason.
Option A — npx skills add (any supported agent)
The README's agent-agnostic path. The CLI auto-detects Claude Code, Cursor, Codex, and similar hosts. Same family of commands as in our npx skills install guide:
# Latest from main — README still labels this "all four"; it pulls the current tree
npx skills add ConardLi/garden-skills
# One skill only (recommended)
npx skills add ConardLi/garden-skills -s web-design-engineer
# Global (~/.skills) instead of per-project
npx skills add ConardLi/garden-skills -s gpt-image-2 --global
# Pin the agent
npx skills add ConardLi/garden-skills -s kb-retriever -a claude-code
Pin a release for CI with a tag-scoped tree URL (example from the README):
npx skills add ConardLi/garden-skills/tree/web-design-engineer-v1.0.0/skills/web-design-engineer
Useful follow-ups from the same CLI: npx skills list, npx skills update, npx skills remove kb-retriever.
Option B — Claude Code plugin marketplace
Inside Claude Code (slash commands, not a website we link):
/plugin marketplace add ConardLi/garden-skills
/plugin install presentation-skills@garden-skills
/plugin install web-design-skills@garden-skills
/plugin install knowledge-base-skills@garden-skills
/plugin install image-generation-skills@garden-skills
| Plugin pack | Skill included |
|---|---|
presentation-skills | web-video-presentation |
web-design-skills | web-design-engineer |
knowledge-base-skills | kb-retriever |
image-generation-skills | gpt-image-2 |
beautiful-article has no row in that table. If you want the fifth skill, use Option A with -s beautiful-article, or copy the folder (Option D).
Option C — pinned .zip from GitHub Releases
Every formal release ships an immutable zip plus SHA-256. Versions advertised in the README "Links" rows as of this writing:
| Skill | README pin |
|---|---|
web-video-presentation | v1.2.2 |
web-design-engineer | v1.3.0 |
gpt-image-2 | v1.0.4 |
kb-retriever | v1.0.1 |
beautiful-article | v0.1.0 |
SKILL=web-design-engineer
VERSION=1.3.0
curl -fsSL -o "${SKILL}.zip" \
"https://github.com/ConardLi/garden-skills/releases/download/${SKILL}-v${VERSION}/${SKILL}-${VERSION}.zip"
curl -fsSL -o "${SKILL}.zip.sha256" \
"https://github.com/ConardLi/garden-skills/releases/download/${SKILL}-v${VERSION}/${SKILL}-${VERSION}.zip.sha256"
shasum -a 256 -c "${SKILL}.zip.sha256"
unzip -q "${SKILL}.zip" -d .claude/skills/
Drop the folder into .agents/skills/ for Cursor or .codex/skills/ for Codex CLI. Checksums matter: a skill is executable instructions plus scripts.
Option D / E — clone, copy, or submodule
git clone https://github.com/ConardLi/garden-skills.git
cp -r garden-skills/skills/web-design-engineer your-project/.claude/skills/
# Cursor / generic agent:
cp -r garden-skills/skills/web-design-engineer your-project/.agents/skills/
Submodule path from the README: git submodule add https://github.com/ConardLi/garden-skills.git vendor/garden-skills, then symlink one skill into .claude/skills/. Checkout a tag such as web-design-engineer-v1.0.0 if you need a frozen SHA.
On claude.ai, the README marks Skills as tested via Settings → Capabilities → Skills. UI labels move — our August 2026 walkthrough of three ways to add skills on Claude used Customize → Skills. Web skills do not automatically appear in Claude Code; filesystem vs Settings are separate.
Which of the five skills should I install first?
Install one. Progressive disclosure still costs description tokens, and these five are all "make something visual or retrieved." Overlap is real.
| Skill | Best first if… | Skip first if… | README version |
|---|---|---|---|
web-design-engineer | You ship pages, dashboards, prototypes, HTML decks | You only need images or a knowledge base | v1.3.0 |
gpt-image-2 | You already have an OpenAI-compatible image API | You only want CSS/HTML in-repo | v1.0.4 |
kb-retriever | You have a knowledge/ tree with indexes | You expected cloud RAG / MCP search | v1.0.1 |
web-video-presentation | You will screen-record a 1920×1080 click-through | You wanted Keynote/PPTX, not Vite | v1.2.2 |
beautiful-article | You will sit through editorial checkpoints | You need a mature, widely battle-tested skill | v0.1.0 |
web-design-engineer — default for most frontend agents
Treats the agent as a design engineer: product context → design system → early v0 → full build. Highlights the README actually ships: a five-dial "Design Read" (variance, motion, density, asset dependence, brand fidelity); extension vs preserve vs overhaul modes; an anti-cliché blocklist; six design schools + 25 anchored style recipes (Linear, Aesop, Pentagram, Bloomberg, Stripe Press, Mid-Century, and more). Browser acceptance runs only when you ask for QA.
If your pain is purple-gradient slop specifically, also read Hallmark (Nutlope) — a different skill with slop-test gates. Garden's take is recipe-and-school based rather than a 57-gate auditor. Pair either with a DESIGN.md so the skill is not inventing brand from scratch.
gpt-image-2 — templates, not a new model
Focused skill for GPT Image 2 and compatible image APIs. Three modes: A Garden local, B host-native delegation, C advisor-only prompt writing. Mode detection runs first so the skill does not silently pick the wrong path. README claims 18 visual categories, 79 structured prompt templates, and a 160+ public case library.
That is a generation workflow skill. If you wanted a giant reverse-engineered prompt catalog instead, our awesome-gpt-image-2 guide covers the 530+ case library. For wiring Claude Code to OpenAI image generation on this site, see generate images from Claude Code.
kb-retriever — local, bounded search
Navigates layered data_structure.md indexes before reading files. Caps exploration at five search rounds. PDF/Excel go through a learn-before-process rule (pdftotext, pdfplumber, pandas). This is not MCP and not a hosted vector DB — it is a skill that teaches the agent to grep and window-read without flooding context. You still need a well-laid-out knowledge/ directory.
web-video-presentation — record-ready Vite decks
Fixed 1920×1080 stage, click/keyboard (chapter, step) cursor, 23 themes, optional TTS (built-in MiniMax mmx-cli + OpenAI TTS via curl; snippets for other providers). Hard checkpoints for script, theme, outline, implementation mode, and audio. Open issues as of this writing include AutoStartGate skipping step 0 when SPACE is pressed (#27 / PR #32) and TTS pronunciation for non-Chinese tokens (#24). Treat "cinematic" as the skill's ambition, not a guarantee on every theme.
beautiful-article — editorial harness, still v0.1.0
Source (URL, PDF, DOCX, Markdown, screenshots) → plan → double-confirmation → build → review → repair. Ten article types with stated retention ratios (longform ~100% down to interactive-explainer ~25% excerpt + 75% rebuild). Eleven theme profiles (tufte, press, bayer, …) as Markdown contracts, not CSS files. Delivery is a self-contained HTML file (optional PDF). Early version: install if you will actually sit at the checkpoints.
How does this compare to rolling your own skills — or to explainx.ai?
| Approach | What you get | Cost |
|---|---|---|
| Garden Skills | Five deep, gallery-backed skills + release zips + Claude Code plugin packs | You adopt ConardLi's taste, checkpoints, and script surface |
Write your own SKILL.md | Exact team procedure, your triggers, your examples | Time; see the first-skill walkthrough |
explainx.ai /skills | Ranked, searchable skills across authors, copy-ready install commands | You still read SKILL.md before production — same as any GitHub pack |
Always-on CLAUDE.md | Rules that apply every session | Context bloat; keep playbooks in skills — CLAUDE.md vs SKILL.md vs MCP |
Garden is a taste pack. It will not replace a registry when you need "the skill this repo's stack actually uses" (Prisma migrate, Vercel deploy, internal review checklist). It will replace a week of writing a design-engineer or image-prompt skill from a blank file.
A sane split: one Garden skill that matches this week's artifact type, plus team skills from /skills or a private .claude/skills/ folder for repo-specific work. Do not paste five visual gardens into every project "just in case."
What should I know before I trust it in production?
README drift. The skills-count badge says 5. The Option A comment still says "Install all four skills." Plugin marketplace JSON (as documented in the README table) omits beautiful-article. Believe the skills/ directory and the per-skill version links, not the leftover "four."
Last push vs star count. Stars were still moving on August 28, 2026. pushed_at on the GitHub API was July 12, 2026. You are installing a popular tree that had not been pushed for about six weeks at publication — pin a release zip if that matters.
Open tracker (honest, not a dump). open_issues_count was 16 (issues + PRs). Visible themes: web-video AutoStartGate (#27), TTS abstraction (#24), a proposal to add another anti-slop design skill (#30), Codex OAuth image mode (PR #25), optional Atlas Cloud for gpt-image-2 (PR #31). The garden is actively forked; it is not frozen polish.
Scripts and supply chain. Skills can include scripts/. Treat npx skills add like installing a package: read SKILL.md, skim scripts, prefer checksummed zips in CI. That is the same discipline we argue in why agent skills are a security risk.
Dependencies you bring. gpt-image-2 needs an image API in Mode A/B. TTS needs a provider. kb-retriever wants grep plus PDF/Excel tooling. web-video-presentation scaffolds Vite + React + TypeScript — you still record the video yourself.
Compatibility (README, not our lab). Marked tested: Claude Code (.claude/skills/ or plugin), Claude.ai web, Cursor (.agents/skills/), Codex CLI (.codex/skills/), Gemini CLI (extension manifest), OpenCode (.opencode/skills/). "Tested" is the maintainer's matrix; re-check after a harness update.
What it is not. Not MCP. Not a general coding curriculum. Not a substitute for the explainx.ai skills registry. Acknowledgments in the README also point at other community lists — we route discovery to /skills and our own guides instead of sending you through a third-party index.
A 15-minute first run
- Install one skill with
-s, not the whole repo. - Restart the agent session so it rescan skills.
- Trigger with a task that matches the frontmatter
description— for design: "Use web-design-engineer to prototype a pricing page; Design Read first, then a v0." - If nothing loads, the description did not match. That is a trigger problem, not a broken install — same debugging as adding skills on Claude.
- For a second skill, pick a non-overlapping category (KB or images), not two design-adjacent packs.
Maintainer checks if you fork: npm run list and npm run validate (Node ≥ 20). The repo package.json is maintainer scripts only; per-skill versions live in each skill's manifest.json.
Related on explainx.ai
- What are agent skills? Complete guide
- npx skills install — Claude Code skills registry
- How to build your first agent skill
- How to add skills to Claude (3 methods)
- Hallmark — anti-AI-slop design skill
- awesome-gpt-image-2 prompt library
- Generate images from Claude Code
- Agent skills security and verification
Official: ConardLi/garden-skills · Releases · Agent Skills spec
Browse ranked skills on explainx.ai: /skills.
Star counts, skill versions, plugin packs, and install commands reflect the public GitHub README, skills/ tree, and GitHub API as of August 28, 2026. The Option A comment still says "four skills" while five exist. Re-read SKILL.md and release checksums before production installs; harness directories and Claude.ai Settings labels change.