Garden Skills

What It Does

Garden Skills is a curated collection of 5 agent skills by ConardLi (Conard Li), one of the most-followed Chinese frontend developers and tech bloggers. The repo targets AI coding agents that support the SKILL.md format (Claude Code, Cursor, Codex, Gemini CLI, OpenCode) and covers creative and productivity use cases: web design, image generation, HTML article creation, video-style web presentations, and local knowledge-base retrieval. Each skill is a self-contained directory with a SKILL.md, references, scripts, and (for two skills) scaffold templates that generate complete Vite+React+TypeScript projects. The collection ships with a Claude Code plugin marketplace configuration, 5 documented install paths, and trilingual README files in English, Chinese, and Japanese. The target users are developers and content creators who want their AI agent to produce polished visual output rather than generic code stubs.

The Good

The gpt-image-2 skill carries 94 structured prompt templates across 18 categories, and the depth is real. The references directory spans 17 subdirectories (ui-mockups, product-visuals, maps, slides-and-visual-docs, poster-and-campaigns, portraits-and-characters, scenes-and-illustrations, editing-workflows, avatars-and-profile, storyboards-and-sequences, grids-and-collages, branding-and-packaging, typography-and-text-layout, assets-and-props, academic-figures, infographics, technical-diagrams). Each template file is a full structured prompt specification, not a stub. For example, references/ui-mockups/live-commerce-ui.md is 227 lines defining a JSON prompt template for generating e-commerce livestream interface mockups, with parameterized fields for host identity, product cards, chat messages, gift popups, and platform-specific styles (Douyin, Xiaohongshu, Taobao). The largest file, references/prompt-writing.md, is 1,005 lines of methodology. This is the largest single-skill structured prompt library GearScope has reviewed.

The 3-mode adaptive runtime in gpt-image-2 is clever engineering. The skill ships a check-mode.js script that probes the environment and returns mode A (Garden local generation with API key), B (host-native, delegate to the agent's own image tools), or C (advisor, produce prompts only). The script reads env vars in priority order: CLI parameters, process.env, .env, .gateway.env, and ~/.gateway.env. The generate.js and edit.js scripts implement OpenAI-compatible API calls with proper multipart form data for edits, b64_json and URL response parsing, and automatic file naming with task slugs and timestamps. When no API key is present, the skill degrades gracefully to a prompt-engineering guide rather than failing silently. This is the kind of fallback design that most image-generation skills describe but few implement.

The web-design-engineer skill has 25 named design style recipes with concrete, paste-ready values. Each recipe file under references/style-recipes/ defines a specific palette, typography stack, spacing system, border-radius strategy, shadow hierarchy, and signature visual moves for a named aesthetic: Linear, Vercel, Raycast, Stripe Press, Apple HIG, Dieter Rams/Braun, Pentagram, Massimo Vignelli, Edward Tufte, Kenya Hara (MUJI), Aesop, Bloomberg Terminal, Bloomberg Businessweek, Monocle, Mailchimp, Headspace, Field.io, Active Theory, Resn, Are.na, Balenciaga, Notion, NYT The Daily, mid-century modern, and Y2K retrofuturism. The INDEX.md organizes them by school (information architecture, editorial, builder SaaS, motion, brutalist, warm humanist) with cross-references. An agent that reads references/style-recipes/linear.md gets exact color values, font names, spacing rules, and 3-4 signature design moves, not vague aesthetic advice.

Both scaffold scripts create real, compiling projects. The beautiful-article/scripts/scaffold.sh creates a Vite+React+TypeScript workspace, installs the reacticle npm package (resolved to v0.2.6), runs npx tsc --noEmit to verify the project compiles, and sets up source/, plan/, review/, article/ directories. The web-video-presentation/scripts/scaffold.sh creates a similar project with 23 built-in themes, each containing its own theme.json and tokens.css. The web-video scaffold also ships a working example chapter (01-example) and a narration extraction pipeline (extract-narrations.ts) that scans all chapter narrations.ts files into a single audio-segments.json for TTS synthesis. Both scaffold scripts were tested and produced working project structures.

The anti-AI-slop philosophy is baked deep into the skills, not bolted on. The web-design-engineer SKILL.md devotes an entire section to "Avoid AI-Style Cliches" with a 7-row table mapping each pattern (purple-pink gradients, rounded card with left-border accent, emoji as icon substitute, Inter as display font) to why it reads as slop and when it is legitimately fine. The reasoning chain is explicit: "AI defaults equal the average of training data, which means all brands averaged together, which means no brand recognized." The web-video-presentation CHAPTER-CRAFT.md includes a "Part 5: Anti-AI-slop patterns" section with detection and repair guidance. This is design discipline encoded as agent instructions, and it works because it explains the why, not just the what.

The Bad

Three of five SKILL.md files are written entirely in Chinese, and so are all 94 image-prompt templates. The gpt-image-2, kb-retriever, and web-video-presentation skills have their core SKILL.md instructions in Chinese. The gpt-image-2 template descriptions (live-commerce UI, product exploded views, neural network architecture diagrams, academic graphical abstracts) are all in Chinese. An English-speaking developer who clones the repo gets a polished trilingual README and two English skills (web-design-engineer, beautiful-article), but the headline draw (94 structured image prompts) is effectively behind a language wall. The web-video-presentation themes are bilingual (theme names and descriptions have both Chinese and English), but the CHAPTER-CRAFT.md, SCRIPT-STYLE.md, and OUTLINE-FORMAT.md references that define how to build chapters are in Chinese. This is the single biggest factor limiting broader adoption.

The primary image-generation mode could not be tested because it requires credentials. Mode A (Garden local generation) needs both ENABLE_GARDEN_IMAGEGEN and OPENAI_API_KEY to be set. Without these, the skill falls back to Mode B or C (prompt advisory only). This means the core value proposition (generate images via GPT Image 2) was verified at the script level (check-mode.js detects the environment correctly, generate.js has proper --help output showing all 12 parameters) but not at the API-call level. The API integration code in generate.js and shared.js appears well-structured (proper env file parsing, multipart form for edits, response parsing with fallback), but no actual image was produced during testing.

The kb-retriever Python script has an undocumented dependency. The scripts/convert_pdf_to_images.py imports pdf2image at module level, but this package is not listed in any requirements.txt, setup.py, or the SKILL.md itself. Running the script on a clean environment produces ModuleNotFoundError: No module named 'pdf2image'. The skill's SKILL.md references the script's existence and instructs the agent to read references/pdf_reading.md before processing PDFs, but neither the SKILL.md nor the reference file mentions that pdf2image must be installed via pip. A developer following the instructions would hit this error immediately.

The beautiful-article scaffold.sh has a minor but real bug at the end of execution. After successfully creating the workspace, installing dependencies, and passing typecheck, the script prints an error: scaffold.sh: line 148: TARGET: unbound variable. The workspace is created correctly and the project compiles, so this is a cosmetic issue in the final success message block (the $TARGET variable reference in the heredoc is not properly expanded). It does not affect functionality, but it is the kind of rough edge that suggests limited testing across environments.

There is no automated test suite for the skills themselves. Unlike DeerFlow (6,776 tests) or CowAgent (222 tests), Garden Skills has no unit tests, integration tests, or CI verification of its scripts. The scaffold scripts are the closest thing to executable verification, but they are not wired into any CI pipeline. A skill author who modifies a template or script has no automated way to check that nothing broke.

Smoke Test Results

Testing was done on a macOS host (aarch64) with Node.js and Python 3 preinstalled. All scripts were run from a shallow clone of the main branch.

Run A: Fresh clone, no deps preinstalled

$ git clone --depth 1 https://github.com/ConardLi/garden-skills.git
$ node skills/gpt-image-2/scripts/check-mode.js
mode: B-or-C recommendation: host-or-advisor garden_mode_enabled: false
OK

$ node skills/gpt-image-2/scripts/check-mode.js --json
{"mode":"B-or-C","recommendation":"host-or-advisor",...}
OK

$ node skills/gpt-image-2/scripts/generate.js --help
Usage: node scripts/generate.js --prompt "..." --image out/otter.png
(12 options listed)
OK

$ bash skills/web-video-presentation/scripts/scaffold.sh --list-themes
(23 themes listed with Chinese names and descriptions)
OK

$ bash skills/beautiful-article/scripts/scaffold.sh --list-themes
(6 themes listed with names, labels, and mood descriptions)
OK

$ python3 skills/kb-retriever/scripts/convert_pdf_to_images.py
ModuleNotFoundError: No module named 'pdf2image'
FAIL

Pass rate: 5 of 6. The only failure is the undocumented pdf2image dependency in the kb-retriever script.

Run B: Clone with Node.js preinstalled

$ cd /tmp/gs-test && bash /tmp/garden-skills/skills/beautiful-article/scripts/scaffold.sh ./test-article --theme=tufte
Reacticle installed: v0.2.6
Typecheck passed
Workspace created with article/, plan/, review/, source/, vite.config.ts, package.json
OK (minor: unbound variable warning on final message line)

$ cd test-article && npx tsc --noEmit
(exit code 0, no errors)
OK

Pass rate: 2 of 2. The scaffold creates a real, typechecking Vite+React+TS project.

Run C: Functional verification

$ node skills/gpt-image-2/scripts/check-mode.js --json | python3 -c "import json,sys; d=json.load(sys.stdin); assert d['mode'] in ['A','A?','B-or-C']; print('mode detection OK')"
mode detection OK

$ node skills/gpt-image-2/scripts/generate.js --help | grep -c "\-\-"
12 (all CLI options present)

$ cat .claude-plugin/marketplace.json | python3 -c "import json,sys; d=json.load(sys.stdin); assert len(d['plugins'])==5; print('plugin marketplace: 5 plugins registered')"
plugin marketplace: 5 plugins registered

$ find skills -name "SKILL.md" | wc -l
5

$ find skills/gpt-image-2/references -name "*.md" | wc -l
94

Functional pass rate: 5 of 5. All structural and script-level checks pass.

What the runs tell you

The scripts work correctly on a clean macOS environment with Node.js preinstalled. The scaffold creates real projects that pass typecheck. The check-mode.js mode detection is accurate and produces valid JSON. The only executable failure is the pdf2image dependency in kb-retriever, and the only untested path is the actual GPT Image 2 API call (Mode A), which requires credentials. The collection is fundamentally sound at the executable level.

Setup Walkthrough

1. Clone the repo or use the skills CLI: npx skills add ConardLi/garden-skills. This auto-detects your agent and installs to the correct directory.
2. For Claude Code users, use the plugin marketplace: /plugin marketplace add ConardLi/garden-skills, then enable individual plugin packs (presentation-skills, web-design-skills, knowledge-base-skills, image-generation-skills, article-skills).
3. To use gpt-image-2 Mode A (actual image generation), set two environment variables: ENABLE_GARDEN_IMAGEGEN=1 and OPENAI_API_KEY=sk-.... Optionally set OPENAI_BASE_URL to point at a compatible gateway and OPENAI_IMAGE_MODEL to use a different model.
4. For the beautiful-article and web-video-presentation skills, run the scaffold script to generate a Vite project: bash /scripts/scaffold.sh ./my-project --theme=. Use --list-themes to see available themes.
5. For kb-retriever PDF processing, install pdf2image manually: pip install pdf2image (also requires poppler on the system: brew install poppler on macOS).

Alternatives

1. anthropics/skills - The official Anthropic skill pack (160,555 stars). Ships a smaller but English-native set of skills including art-design-builder and web-app-builder. Prefer it if you want official, English-language skills with broad compatibility.
2. addyosmani/agent-skills (77,559 stars) - A large English-language collection of development-focused skills. Better for engineering workflows; does not match Garden Skills on creative/design depth.
3. cgallic/kai-cmo-harness - A marketing-focused skill pack. Covers SEO, content, and ads but lacks the design-system depth of web-design-engineer. Prefer it for pure marketing automation.