现在用 AI 做项目,经常会碰到这种尴尬的情况:
写代码,它喜欢乱造轮子,能用十行解决的事,非要写一大坨。
做页面,它能把功能搭出来,但审美一言难尽。
抓数据,它很多网页抓不全、抓不准,甚至直接没权限。
不是不会做,而是做得太糙了,完全不注意工程细节。
针对上面这些问题,我整理了三个开源项目,帮助你的 AI 更准确地实现需求。
一、少写代码
这个项目,目前在 GitHub 上已经 6.8 万多 Star 了,能让你的 AI 少写一半的代码。
它叫 Ponytail,是一个可以放进 Agent 里的插件。
项目图标是一个扎着马尾辫、戴着眼镜的资深程序员形象,意为此项目对修改屎山代码很有经验。
Ponytail 的思路不是事后帮你压缩代码,而是在 AI 动手之前,先给它套上一层“少写代码”的纪律。
你只是告诉 AI 代码要写短一点。
Ponytail 则是让它在动手前先作判断:
-
这个功能是不是真的需要存在;
-
项目里是不是已经有类似实现;
-
标准库能不能解决;
-
浏览器或平台原生能力能不能覆盖;
.....
实在不行,再写最小可用实现。
它不是让 AI 为了短而短,而是让 AI 先从“不写代码”开始思考。
比如做一个日期选择器,能力弱一点的 Agent 可能会装库、写 wrapper、加样式、处理一堆边界情况。
Ponytail 会先问 Agent 浏览器不是已经有 吗?
然后直接原生能力解决,决不会多写一行代码。
在 Agent 不同运行阶段,它都会自动把规则塞进去。
比如,在会话刚开始,它会通过 SessionStart 加载规则。
如果主 Agent 又启动了子 Agent,它还会通过 SubagentStart 把同一套规则继续传给子 Agent,避免子 Agent 乱造轮子的情况。
这就很舒服,因为它贯穿到整个 Agent 工作流里,不用你一直操心。
作者在一个真实的 FastAPI + React 开源项目里,用 ponytail 作了实测,模型用的是 Haiku 4.5。
在前端这种最容易过度工程化的任务里,Ponytail 的效果非常强。
-
date picker,baseline 平均新增 404 行代码,Ponytail 只新增 23 行。
-
color picker ,baseline 是 287 行,Ponytail 还是 23 行。
-
file dropzone 任务里,baseline 是 251 行,Ponytail 降到 95 行。
当然,它也不是所有任务都能砍很多。
像后端 CRUD 这种本来就没太多水分的任务,各组差距就不大。
最后汇总下来,在 12 个真实任务里,Ponytail 相比普通 Claude Code,平均减少了 54% 的新增代码行数,token 降了 22%,成本降了 20%,耗时也少了 27%。
安全性方面,Ponytail 的通过率是 100%。
插件支持 Claude Code、Codex、OpenCode 等主流 AI 编程工具。
安装到 Claude Code ,只需执行下面两行命令即可安装好。
/pluginmarketplaceaddDietrichGebert/ponytail/plugininstallponytail@ponytail
另外插件还提供了几个模式,lite 轻量、full 完整、ultra 激进,不想用了直接 off 关掉。
项目链接:
https://github.com/DietrichGebert/ponytail
二、克隆网页
这个项目只需要一行命令,就可以1:1复刻网页。
普通 AI 克隆网页,很多时候只是看截图,然后凭感觉猜颜色、字体、间距和布局。所以做出来的页面远看像,近看全是问题。
而这个项目是一个网页逆向工程系统,会把浏览器里的每个细节都尽量转化成可复现的前端代码。
它的工作流是这样的:
首先,它会先让 AI 通过浏览器真正打开目标网站,对页面做一轮全域采集,通过 getComputedStyle() 直接读取浏览器里的真实计算值。
比如网页里具体的颜色的数值是多少,字体是什么,间距、圆角、阴影具体参数是什么。
拿到这些设计 Token 之后,AI 才会开始先搭底座,全局 CSS、Tailwind 配置、静态资源都会先对齐原站。
接下来,它会把整个页面拆成一个个组件。
每个组件都会先生成一份规格书,写清楚这个区块的结构、响应式变化和交互状态。
之后,它再让多个 builder agent 并行施工。每个 Agent 负责一个组件,并且放在独立的 git worktree 里开发。
这样导航栏、Hero 区、页脚可以同时做,互不干扰。
最后,AI 把这些组件合并回主项目,拼成完整页面。
下面是作者给到的实测视频:
这比让一个 AI 从头到尾写完整页面稳定很多。
因为前端页面本来就是模块化的,拆成区块以后,Agent 更容易对齐原站,也更容易做视觉校正。
当然,这个项目更适合用在你自己的网站迁移、学习优秀页面结构这些场景。
不适合拿来直接搬别人的商业网站上线。
安装方面,这个项目官方不建议直接 clone 原仓库,而是先在 GitHub 页面点击 Use this template,创建一份属于你自己的项目。
然后把你自己的仓库 clone 到本地:
gitclonehttps://github.com/YOUR-USERNAME/YOUR-NEW-REPOSITORY.gitcdYOUR-NEW-REPOSITORY
接着启动 Claude Code,输入:
/clone-website<目标网站URL>
接下来,AI 就会打开目标网页,采集页面结构、提取设计 Token 和资源,然后开始重建这个网站。
项目要求 Node.js 24+,官方推荐 Claude Code,但 Codex CLI、Cursor、Windsurf、Gemini CLI 等 AI 编程工具也支持。
项目链接:
https://github.com/JCodesMore/ai-website-cloner-template
三、免费爬可靠数据
这个项目能把任何网页,变成 AI 能直接读取的结构化数据。
你给它一个 URL,它可以输出:
-
干净的 Markdown,去掉导航栏、广告、页脚这些杂乱内容;
-
结构化 JSON,你定义字段,它按 schema 提取;
-
也可以返回截图、HTML 和网页元数据。
不只是抓单页,它还能爬整个网站、解析本地文件、做 arXiv 论文搜索、搜索 GitHub 仓库信息。
而且, Firecrawl 主推无 Key 模式,不用配环境、申请密钥,直接就能用。
Firecrawl 的能力有三层:
-
你可以给它一个 URL,抓单页;也可以给它一个网站入口,爬整站;如果不知道具体页面在哪,还可以先用
map把站内 URL 列出来。 -
然后,它会处理网页抓取、JS 渲染、点击、滚动、等待这些动作,再把网页里的正文、图片、链接和元数据提取出来,转成 AI 更容易读取的 Markdown、JSON、HTML 或截图。
-
而且,你不一定要提前准备好 URL,只要描述需求,Firecrawl Agent 就能自己搜索、打开网页、导航、抓取,并把结果和来源返回。
以前 Agent 读网页,要自己处理搜索和解析。
现在 Firecrawl 先把这些脏活做掉,最后交给 AI 的就是干净的数据。
这个项目已经不是一个小众工具了。
它是 GitHub 全站 Top 100 级别的开源仓库,目前已经有 130K+ Star。
官方数据显示,有 15 万多家公司在用 Firecrawl,其中包括 Apple、Canva、Lovable、Stanford、Zapier、Replit 这些团队。
安装 Firecrawl 只需一行命令:
npx-yfirecrawl-cli@latestinit--all--browser
它支持 Claude Code、OpenCode、Cursor、Windsurf、Codex CLI、Gemini CLI、Antigravity。
项目链接:
https://github.com/firecrawl
少写没必要的代码,做出更准的页面,拿到更干净的数据,AI 做项目的可用性就会高很多。
这期就先分享这三个项目,都是我觉得真正能改善 AI 工作流的开源项目。
我们下期再见!
·················END·················
When using AI for projects these days, you often run into this awkward situation:
When it writes code, it loves reinventing the wheel. What could be done in ten lines, it insists on writing a huge pile of code.
When it builds pages, it can get the functionality working, but the aesthetics are hard to put into words.
When it scrapes data, it often fails to capture everything, gets things wrong, or even has no access at all.
It's not that it can't do it, it's that the work is too rough and it completely ignores engineering details.
To address the problems above, I've put together three open-source projects to help your AI implement your requirements more accurately.
1. Write Less Code
This project has already earned over 68,000 stars on GitHub. It can help your AI write half as much code.
It's called Ponytail, and it's a plugin you can put into an Agent.
The project icon is a senior programmer with a ponytail and glasses, symbolizing that this project is experienced in fixing messy legacy code.
Ponytail's approach isn't to compress your code after the fact. Instead, before the AI even starts, it wraps the AI in a "write less code" discipline.
You just tell the AI to write shorter code.
Ponytail makes the AI make a judgment before starting:
Does this feature actually need to exist?
Is there already a similar implementation in the project?
Can the standard library solve it?
Can native browser or platform capabilities cover it?
.....
If none of those work, then and only then write the smallest viable implementation.
It doesn't make the AI write short for the sake of being short. It makes the AI start thinking from "don't write code" first.
For example, when building a date picker, a less capable Agent might install libraries, write wrappers, add styles, and handle a bunch of edge cases.
Ponytail first asks the Agent, doesn't the browser already have ?
Then it just uses the native capability, without writing a single extra line of code.
At different stages of the Agent's run, it automatically injects the rules.
For example, at the start of a session, it loads the rules via SessionStart.
If the main Agent launches a sub-agent, it also passes the same rules to the sub-agent through SubagentStart, preventing the sub-agent from reinventing the wheel.
This feels great because it runs through the entire Agent workflow, so you don't have to keep worrying about it.
The author tested Ponytail on a real FastAPI + React open-source project, using Haiku 4.5 as the model.
On front-end tasks, which are the easiest to over-engineer, Ponytail's effect is very strong.
For the date picker, the baseline added an average of 404 lines of code, while Ponytail only added 23 lines.
For the color picker, the baseline was 287 lines, and Ponytail was still 23 lines.
For the file dropzone task, the baseline was 251 lines, and Ponytail reduced it to 95 lines.
Of course, it can't cut a lot on every task.
For tasks like back-end CRUD, which don't have much fluff to begin with, the differences between groups are small.
In the final summary across 12 real tasks, compared with regular Claude Code, Ponytail reduced the average number of new lines of code by 54%, tokens by 22%, cost by 20%, and time by 27%.
In terms of safety, Ponytail's pass rate is 100%.
The plugin supports mainstream AI coding tools such as Claude Code, Codex, and OpenCode.
To install it into Claude Code, you just run the two commands below.
The plugin also provides several modes: lite, full, and ultra. If you don't want to use it, you can simply turn it off.
Project link:
https://github.com/DietrichGebert/ponytail
2. Clone Web Pages
This project can replicate a webpage 1:1 with just one command.
When ordinary AI clones a webpage, it often just looks at a screenshot and then guesses colors, fonts, spacing, and layout based on intuition. So the resulting page looks similar from afar, but up close it's all problems.
This project, however, is a web reverse-engineering system. It converts every detail in the browser into reproducible front-end code as much as possible.
Its workflow goes like this:
First, it has the AI actually open the target website in a browser and perform a full-coverage scan of the page, using getComputedStyle() to directly read the real computed values in the browser.
For example, the exact color values in the page, what the fonts are, and the specific parameters for spacing, border radius, and shadows.
After obtaining these design tokens, the AI starts by building the foundation first, aligning global CSS, Tailwind config, and static assets with the original site.
Next, it breaks the entire page into individual components.
Each component first gets a spec document that clearly describes the structure, responsive changes, and interaction states of that section.
Then, it has multiple builder agents work in parallel. Each agent is responsible for one component and develops it in a separate git worktree.
This way, the navigation bar, Hero section, and footer can be built at the same time without interfering with each other.
Finally, the AI merges these components back into the main project to assemble the complete page.
Below is the actual test video provided by the author:
This is much more stable than having one AI write the entire page from start to finish.
Because front-end pages are modular by nature, once they are split into sections, it's easier for Agents to align with the original site and do visual corrections.
Of course, this project is better suited for scenarios like migrating your own website or learning from excellent page structures.
It is not suitable for directly copying someone else's commercial website and putting it into production.
For installation, the project officially recommends not cloning the original repository directly. Instead, click "Use this template" on the GitHub page to create your own project.
Then clone your own repository locally:
Next, launch Claude Code and enter:
The AI will then open the target webpage, collect the page structure, extract design tokens and resources, and begin rebuilding the website.
The project requires Node.js 24+. Claude Code is officially recommended, but AI coding tools like Codex CLI, Cursor, Windsurf, and Gemini CLI are also supported.
Project link:
https://github.com/JCodesMore/ai-website-cloner-template
3. Scrape Reliable Data for Free
This project can turn any webpage into structured data that AI can read directly.
Give it a URL, and it can output:
Clean Markdown with navigation bars, ads, footers, and other cluttered content removed;
Structured JSON, where you define the fields and it extracts according to the schema;
It can also return screenshots, HTML, and webpage metadata.
It doesn't just scrape single pages. It can also crawl entire websites, parse local files, search arXiv papers, and search GitHub repository information.
Moreover, Firecrawl promotes a keyless mode. You don't need to configure an environment, apply for an API key, or do anything like that. You can just use it directly.
Firecrawl has three layers of capability:
You can give it a URL to scrape a single page; you can give it a website entry point to crawl the whole site; and if you don't know where a specific page is, you can first use map to list all the URLs within the site.
Then, it handles web scraping, JS rendering, clicking, scrolling, and waiting. After that, it extracts the main text, images, links, and metadata from the page, converting them into Markdown, JSON, HTML, or screenshots that are easier for AI to read.
Moreover, you don't necessarily need to prepare a URL in advance. Just describe what you need, and the Firecrawl Agent can search, open webpages, navigate, scrape, and return both the results and the sources on its own.
In the past, when an Agent read a webpage, it had to handle searching and parsing itself.
Now, Firecrawl does all that dirty work first, and what's handed to the AI is clean data.
This project is no longer a niche tool.
It's an open-source repository at the GitHub Top 100 level, with over 130K stars right now.
Official data shows that more than 150,000 companies are using Firecrawl, including teams like Apple, Canva, Lovable, Stanford, Zapier, and Replit.
Installing Firecrawl only takes one command:
It supports Claude Code, OpenCode, Cursor, Windsurf, Codex CLI, Gemini CLI, and Antigravity.
Project link:
https://github.com/firecrawl
Write less unnecessary code, produce more accurate pages, and get cleaner data. That makes AI much more usable for real projects.
That's all for this issue. These three projects are ones I genuinely believe can improve AI workflows.
See you next time!