花园开源了 4 个神级 Skills
做视频 · 写网页 · 生成图 · 查知识库一篇讲完怎么用
Skill 教程 + 开源推荐 + 实操指南技术编辑部/ 工具评测 / 阅读 8 min◎ 导 语ConardLi(花园)在 GitHub 一次性放出 4 个 Agent Skills,MIT 协议、可商用。四个方向:把文章做成可录屏的"伪视频"、让 AI 写出不像 AI 做的网页、用结构化 prompt 出图、本地知识库带溯源问答。下面先看每个 Skill 能干什么,再讲怎么装。
| 01 | / 4 个 Skills 速查 |
1. web-video-presentation:把文章变成可录屏的"伪视频"
做什么:把你写好的脚本、文章或者产品介绍拆成一小段一小段的 narration,每段对应一屏 16:9 的画面,键盘点一下就往前推,录屏出来就是视频。
适用场景:技术分享、产品发布、知识付费、不想碰 Premiere 但又想要"视频感"的人。
硬卖点:
1920×1080 固定舞台——录屏不会跳尺寸,自动适配任意屏幕;
23 个内置主题,editorial、terminal、engineering、Swiss International 每个风格都不一样;
5 个硬协作 checkpoint——脚本、主题、大纲、实现模式、音频,每一步都卡住等你确认;TTS 引擎可以换,内置 MiniMax 和 OpenAI,也预留了 ElevenLabs、edge-tts、Azure、Google Cloud、macOS say 的接口。
23 个主题听起来多,实际大部分人可能 editorial 和 terminal 两个就够了。TTS 那边 MiniMax 和 OpenAI 直接能用,换其他引擎要自己配。
先装的人:以录屏出片为目标的创作者。
01npx skills add ConardLi/garden-skills -s web-video-presentation
2. web-design-engineer:让 AI 写出去掉 AI 味的网页
做什么:把 Agent 从"生成能跑的代码"拉到"做出有品味的前端"。落地页、Dashboard、原型、幻灯片、动画——前端能见的它基本都管。
适用场景:被吐槽"一眼 AI 拼的网页"的开发者、做独立产品的全栈、需要快速出 v0 的前端。
硬卖点:6 步工作流——需求→上下文→设计系统→v0→完整构建→验证,每一步有明确的产出;6 大设计流派 + 25 个风格菜谱,Linear、Aesop、Pentagram、Bloomberg、Stripe Press、Mid-Century 都有,每个菜谱直接给了配色、字体、签名手法和反例清单;还带了一个反 AI 套路的 blocklist,通用 AI UI 习气在 v0 阶段就会被拦下来。
25 个菜谱理论上都能用,但试下来 Linear、Aesop、Mid-Century 这几个主流风格效果最稳。偏实验的流派落地门槛不低。
先装的人:要做产品页、SaaS Landing、Dashboard 的。
01npx skills add ConardLi/garden-skills -s web-design-engineer
3. gpt-image-2:我拿来写 prompt,比自己憋的快三倍
实话实说,这 Skill 我装的时候以为就是个提示词收藏夹——80 多个模板塞在 references/ 里,谁不会翻啊?
用了一周发现我想错了。它真正值钱的地方不是模板多,是它知道什么时候用哪个模板。你扔一句「帮我出张信息图」,它会先判断你有没有图像 API、再选分类、再套结构——而不是一股脑把 80 个模板全塞进上下文。
我上次做技术架构图,自己写 prompt 憋了十分钟出来的东西还是像「科技峰会背景板」。换它走了一遍流程,出来的图直接能用。
三个小毛病也得说:
· 必须有 API key。Mode C(纯文字顾问)能帮你写 prompt,但拿不到图。
· 模板太多反而不好翻。80 个模板放在 references/ 里,你如果不知道自己要什么,得一个个翻目录。
· 图像编辑那条流水线我还没跑通,成功率比生成低不少。
实际测试:
- 让它生成一张微信公众号16:9封面提示词,以及一张3:4小红书风格信息图描述词
来看看实际文生图的效果:
再来一张小红书风格信息图:
先装的人:手头有 OpenAI 兼容图像 API、而且烦透了跟 Midjourney 反复斗 prompt 的人。
01npx skills add ConardLi/garden-skills -s gpt-image-2
4. kb-retriever:在本地知识库里做问答
做什么:给你本地的 knowledge/ 目录加一个问答入口。能读 Markdown、纯文本、PDF、Excel,不靠把整份文件塞进上下文。
适用场景:本地有结构化文档库、要从 PDF 或 Excel 里抽证据、不想被大文件撑爆上下文窗口。
硬卖点:先用 data_structure.md 索引文件导航,找到目标再搜,避免一上来就加载整份文件;碰到 PDF 和 Excel 会先学处理方法再动手抽数据;检索最多跑 5 轮,不会无限试探下去;工具链用了 grep、pdftotext、pdfplumber 加 pandas,答案会标注来源。不过我没在超大型知识库上跑过,上千个文件不知道会不会卡。
先装的人:本地有知识库、需要带溯源回答问题的人。
01npx skills add ConardLi/garden-skills -s kb-retriever
版本(以 GitHub Releases 为准):gpt-image-2 v1.0.3、kb-retriever v1.0.0。web-video-presentation 和 web-design-engineer 的版本见 Releases 页。兼容环境见第四节。
| 02 | / 5 种安装方式,该选哪一种 |
仓库支持 5 种装法,按场景对号入座:
| # | 方式 | 入口 | 适合谁 |
| A | skills CLI(npx) | npx skills add ConardLi/garden-skills | 默认推荐。自动识别 Agent 并落到对应目录 |
| B | Claude Code 插件市场 | /plugin marketplace add | 只用 Claude Code、习惯插件管理 |
| C | Releases 的 .zip + SHA-256 | GitHub Releases 下载 | CI、内网、需要锁定版本 |
| D | git clone 后手动复制 | cp -r skills/ | 想 fork 自己改 |
| E | git submodule + tag pin | git submodule add … | 大型工程要 vendor 并跟随上游 |
三个常用 flag:-s
| ⚡ 建议 | 拿不准的话先用 A。 大部分场景下 CLI 已经够用,CI 和生产环境再换到 C 做版本固定。 |
| 03 | / 实操:一行命令装 web-design-engineer |
走一遍最小可复现路径:
1 终端运行安装命令
01npx skills add ConardLi/garden-skills -s web-design-engineer
CLI 会自动识别你当前的 Agent 环境,把 Skill 放到对应目录。不过第一次跑会慢一些,它要从 GitHub 拉包。
2 确认落点
不同 Agent 目录不一样:
Claude Code → .claude/skills/web-design-engineer/
Cursor → .agents/skills/web-design-engineer/
Codex CLI → .codex/skills/web-design-engineer/
3 验证 Skill 已激活
在 AI 里说一句话:
用 web-design-engineer 帮我做一个 SaaS Landing 的 design system
如果 AI 主动列出六步流程、开始列 design tokens(颜色、字体、间距、动效),说明 Skill 已被加载。
4 生产环境固定到 tag
别用 main 分支,用 tag 锁版本:
01npx skills add ConardLi/garden-skills/tree/web-design-engineer-v1.1.0/skills/web-design-engineer
这样每次拉到的字节永远一致。
实际可以使用自然语言安装
| 04 | / 6 种 Agent 兼容性 |
据仓库文档,以下环境已测试通过:
| Agent / 运行时 | Skill 安装目录 |
| Claude Code | .claude/skills/ |
| Claude.ai(Web) | Settings → Capabilities → Skills |
| Cursor | .agents/skills/ |
| Codex CLI | .codex/skills/ |
| Gemini CLI | extension manifest |
| OpenCode | .opencode/skills/ |
Skill 格式本身是通用的——你的 Agent 只要支持 Skills 规范,把文件夹拷过去就能用。
| 05 | / 按场景先装哪一个 |
仓库地址:github.com/ConardLi/garden-skills,MIT,可商用、可 fork。(注意:公众号会剥除外链,建议访问仓库查看。)
按你手头的活儿对号入座:
▸ 做技术分享 / 课程 / 产品 demo 视频 → web-video-presentation
▸ 做产品页 / Landing / Dashboard → web-design-engineer
▸ 量产配图 / UI 稿 / 信息图 → gpt-image-2
▸ 本地有知识库要带溯源问答 → kb-retriever
如果还拿不准,直接去仓库 README 确认:github.com/ConardLi/garden-skills。
Garden Open-Sources 4 God-Tier Skills
Make Videos · Write Webpages · Generate Images · Query Knowledge Base—All Explained in One Article
ConardLi (Garden) has released 4 Agent Skills on GitHub at once, under the MIT license, ready for commercial use. Four directions: turning articles into screen-recordable "pseudo-videos", making AI write webpages that don't look AI-generated, generating images with structured prompts, and local knowledge base Q&A with source tracing. Below, we first look at what each Skill can do, then explain how to install them.
1. web-video-presentation: Turn Articles into Screen-Recordable "Pseudo-Videos"
What it does: Break your script, article, or product introduction into small narration segments, each corresponding to a 16:9 screen. Press a key to advance, and the screen recording becomes a video.
Suitable scenarios: Tech sharing, product launches, paid knowledge, or anyone who wants a "video feel" without touching Premiere.
Hard selling points:
1920×1080 fixed stage—screen recording won't jump sizes, automatically adapts to any screen;
23 built-in themes—editorial, terminal, engineering, Swiss International, each with a distinct style;
5 hard collaboration checkpoints—script, theme, outline, implementation mode, audio, each step waits for your confirmation; TTS engine is swappable, built-in MiniMax and OpenAI, with interfaces reserved for ElevenLabs, edge-tts, Azure, Google Cloud, and macOS say.
23 themes sound like a lot, but in reality, most people will likely only need editorial and terminal. For TTS, MiniMax and OpenAI work directly; switching to other engines requires your own configuration.
Install first if: You are a creator aiming to produce videos via screen recording.
01npx skills add ConardLi/garden-skills -s web-video-presentation
2. web-design-engineer: Make AI Write Webpages Without the AI Smell
What it does: Pull the Agent from "generating runnable code" to "creating tasteful front-end". Landing pages, Dashboards, prototypes, slides, animations—basically anything visible in the front-end is covered.
Suitable scenarios: Developers criticized for "obvious AI-assembled webpages", full-stack developers of independent products, front-end developers needing a quick v0.
Hard selling points: 6-step workflow—requirements → context → design system → v0 → full build → verification, with clear outputs at each step; 6 major design schools + 25 style recipes, including Linear, Aesop, Pentagram, Bloomberg, Stripe Press, Mid-Century, each recipe directly providing color palettes, fonts, signature techniques, and anti-pattern lists; also includes a blocklist against AI clichés, stopping generic AI UI habits at the v0 stage.
Theoretically all 25 recipes are usable, but in practice, Linear, Aesop, and Mid-Century are the most stable mainstream styles. Experimental schools have a higher threshold for implementation.
Install first if: You need to build product pages, SaaS Landings, or Dashboards.
01npx skills add ConardLi/garden-skills -s web-design-engineer
3. gpt-image-2: I Use It to Write Prompts, Three Times Faster Than My Own
To be honest, when I installed this Skill, I thought it was just a prompt collection—over 80 templates stuffed in references/, who can't just browse them?
After using it for a week, I realized I was wrong. Its real value isn't the number of templates, but knowing which template to use when. You throw in a sentence like "Help me generate an infographic", it first checks if you have an image API, then selects the category, then applies the structure—instead of dumping all 80 templates into the context.
Last time I made a technical architecture diagram, I spent ten minutes writing a prompt and the result still looked like a "tech summit backdrop". Switching to it and going through the process, the resulting image was ready to use.
Three minor flaws to mention:
· Must have an API key. Mode C (pure text consultant) can help you write prompts, but you won't get the image.
· Too many templates make them hard to browse. With 80 templates in references/, if you don't know what you want, you have to flip through the directory one by one.
· I haven't fully run through the image editing pipeline yet; its success rate is much lower than generation.
Actual test:
- Asked it to generate a 16:9 cover prompt for a WeChat Official Account, and a 3:4 Xiaohongshu-style infographic description
Let's see the actual text-to-image results:
Another Xiaohongshu-style infographic:
Install first if: You have an OpenAI-compatible image API and are tired of fighting prompts with Midjourney.
01npx skills add ConardLi/garden-skills -s gpt-image-2
4. kb-retriever: Q&A in Your Local Knowledge Base
What it does: Adds a Q&A entry to your local knowledge/ directory. Can read Markdown, plain text, PDF, Excel, without stuffing the entire file into the context.
Suitable scenarios: You have a structured local document library, need to extract evidence from PDFs or Excel files, or don't want large files to blow up the context window.
Hard selling points: First uses data_structure.md to index and navigate files, finds the target before searching, avoiding loading the entire file upfront; encounters PDFs and Excel files, it learns the processing method before extracting data; retrieval runs at most 5 rounds, won't probe infinitely; toolchain uses grep, pdftotext, pdfplumber plus pandas, and answers are annotated with sources. However, I haven't tested it on ultra-large knowledge bases, so I don't know if it will lag with thousands of files.
Install first if: You have a local knowledge base and need Q&A with source tracing.
01npx skills add ConardLi/garden-skills -s kb-retriever
Versions (subject to GitHub Releases): gpt-image-2 v1.0.3, kb-retriever v1.0.0. See the Releases page for web-video-presentation and web-design-engineer versions. Compatibility environment is in Section 4.
The repository supports 5 installation methods, choose based on your scenario:
Three common flags: -s
Follow the minimal reproducible path:
1 Run the installation command in the terminal
01npx skills add ConardLi/garden-skills -s web-design-engineer
The CLI will automatically detect your current Agent environment and place the Skill in the corresponding directory. However, the first run will be slower as it pulls the package from GitHub.
2 Confirm the drop point
Different Agents have different directories:
Claude Code → .claude/skills/web-design-engineer/
Cursor → .agents/skills/web-design-engineer/
Codex CLI → .codex/skills/web-design-engineer/
3 Verify the Skill is activated
Say a sentence in the AI:
Use web-design-engineer to help me create a design system for a SaaS Landing page
If the AI actively lists the six-step process and starts listing design tokens (colors, fonts, spacing, animations), it means the Skill has been loaded.
4 Fix to a tag in the production environment
Don't use the main branch, use a tag to lock the version:
01npx skills add ConardLi/garden-skills/tree/web-design-engineer-v1.1.0/skills/web-design-engineer
This ensures the bytes pulled are always consistent.
Actual usage allows natural language installation
According to the repository documentation, the following environments have been tested and passed:
The Skill format itself is universal—as long as your Agent supports the Skills specification, copying the folder over makes it work.
Repository address: github.com/ConardLi/garden-skills, MIT, commercial use allowed, forking allowed. (Note: The official account strips external links, so it is recommended to visit the repository directly.)
Match based on your current work:
▸ Making tech sharing / courses / product demo videos → web-video-presentation
▸ Making product pages / Landings / Dashboards → web-design-engineer
▸ Mass-producing illustrations / UI drafts / infographics → gpt-image-2
▸ Local knowledge base requiring Q&A with source tracing → kb-retriever
If you are still unsure, go directly to the repository README to confirm: github.com/ConardLi/garden-skills.








