5 GitHub Repos: Kill AI Slop, Go Viral, Make Money

频道: Greg Isenberg
视频: https://www.youtube.com/watch?v=9_SZFIW7tus
原文语言: en
统计: 共 22 轮 · Greg 21

[0:00] Greg

I don't know about you, but I've been spending a lot more time on GitHub lately, and there's a real reason behind it. I mean, it's turned into one of the best places to get an unfair advantage in the agentic era because you're seeing the tools people will be talking about in 6 months today. A lot of the things that eventually become SaaS companies, agencies, workflows, and startup ideas show up there first. It's really like that famous quote, you know, the future is already here, it just isn't evenly distributed yet. So, I'm trying something new with today's episode. I went through a ton of GitHub repos that have been getting attention over the last 30 days, and I picked the six that I think you should know about. If you're trying to build products, if you're trying to make money, if you're trying to save time, or you just want to get more done with AI. Now, some are more technical than others, but I'll explain them all in plain English, you know, what each one does, why it matters, how to actually use it, and how you can try it for yourself today. The last one is from Nvidia, a couple of the other ones are from smaller devs, but every one of them is something I can picture someone using to make money or to save real time.

不知道你们怎么样，反正我最近在 GitHub 上泡的时间多了不少，而且这背后是有实实在在的原因的。说真的，在这个 agent 时代，GitHub 已经成了抢占不公平优势的最佳去处之一，因为你今天在上面看到的工具，就是半年后大家都在热议的东西。很多最后长成 SaaS 公司、agency 服务商、工作流和创业点子的东西，都是先在那里冒头的。这真的就像那句名言说的：未来已经来了，只是分布得还不均匀。所以今天这期我想尝试点新东西。我翻了一大堆过去 30 天里受到关注的 GitHub 仓库，从中挑出了我觉得你应该知道的六个——不管你是想做产品、想赚钱、想省时间，还是只想借助 AI 多干点事。当然，其中有些比别的更偏技术，但我会全部用大白话讲清楚：每个是干什么的、为什么重要、具体怎么用，以及你今天怎么就能亲手试一试。最后一个来自 NVIDIA，另外有几个出自规模较小的开发者，但每一个我都能想象到有人拿它来赚钱，或者实打实地省下时间。

[1:11] Greg

All six are free and open source, meaning the code is free to use, but as a lot of you know, some have dependencies like one has a dependency on 11 Labs, which does cost money, so be aware of that. Uh let's get into the episode. I'll see you at the end, and if you haven't liked and commented and subscribed, just do it to get your algorithm pumping [music] with a lot of valuable content like this. Let's go.

这六个全都免费开源，也就是说代码可以免费使用。不过很多人也知道，有些会有依赖项，比如其中一个就依赖 ElevenLabs，那是要花钱的，这点要留意。好，我们进入正题。节目最后见。如果你还没点赞、评论、订阅，赶紧去点一下，让你的算法源源不断地给你推送这类有价值的内容。我们开始吧。

[1:37]

The startup by this podcast is sleeping time, baby. [music]

（片头插播）The Startup Ideas Podcast，睡觉时间到了，宝贝。（音乐）

[1:45] Greg

So, the first repo is my friend Peter Yang's No AI Slop Skill, and I like this one because everyone who uses AI for writing has felt this exact problem. I mean, you ask ChatGPT or Claude to help you write something, it's a tweet, it's a LinkedIn post, and technically the writing is fine, the grammar's fine, the syntax is fine, but it has this like weird smell to it. Maybe it sounds like too polished or too symmetrical or formulaic. Um sometimes it has like it's not X, but it's Y. Uh you know, or or uses the word quietly a lot. Um you know, I think I feel like it feels like uh reading a keynote from a fake SaaS conference. And once you notice it, I I feel like you can't unsee it. You know, a lot of people are able to smell it now. So, what's really cool is Peter's skill is basically an editor that strip strips those patterns out while trying to preserve your actual voice. Um the last part's important because a lot of the writing tools make your writing cleaner, but sand off the interesting parts and make everyone sound the same. And if you're going to sound the same, it's hard to stand out.

第一个仓库是我朋友 Peter Yang 做的 No AI Slop Skill。我喜欢它，是因为每个用 AI 写东西的人都碰到过这个问题。比如你让 ChatGPT 或 Claude 帮你写点东西，可能是一条推文，也可能是一篇 LinkedIn 帖子，从技术上讲写得没毛病，语法没问题，句法也没问题，但它就是有股说不上来的怪味儿。可能是读起来太精雕细琢、太工整对称，或者太套路化。有时候会冒出那种「不是 X，而是 Y」的句式，或者特别爱用 quietly（悄然）这个词。我感觉读起来就像在看一场山寨 SaaS 大会上的主题演讲。而且一旦你注意到了，就再也没法视而不见。现在很多人都能闻出这股味儿了。所以很酷的一点是，Peter 这个 skill 本质上就是一个编辑，它会把这些套路剔掉，同时尽量保留你本来的语气和风格。最后这一点很重要，因为很多写作工具能让你的文字更干净，却把有意思的部分也一并磨平了，让所有人听起来都一个样。而如果你听起来跟别人都一样，就很难脱颖而出。

[2:57] Greg

Uh you know, you you actually want to do the opposite. Um so, what's cool about uh this this uh repo is it hunts for the patterns that make writing feel AI-generated and helps the draft feel like a real human being wrote it. Uh why does that matter? Like I said, if you're building products today, you're writing constantly and you want to stand out. You know, I I personally am writing tweets, I'm writing landing pages, I'm writing cold emails, uh launch posts, product updates, onboarding copy. If you have investors, you're writing investor updates. And you don't want to sound like generic AI because you you know, people are going to feel it. Uh they might not say this was written by AI. Like they might not respond and say it, uh but they'll just trust you less. Um they're just going to keep scrolling, they're not going to reply. Uh they might even believe what you're saying because they're like this is AI-written.

你其实应该反着来。所以这个仓库很酷的地方在于，它会专门揪出那些让文字显得像 AI 生成的套路，帮你把草稿改得像是真人写的。为什么这很重要？就像我刚才说的，如果你现在在做产品，你就得一直写东西，而且你希望自己能脱颖而出。拿我自己来说，我要写推文、写落地页、写陌生开发信，还有产品发布帖、产品更新、新用户引导文案。如果你有投资人，还得写投资人更新。你肯定不想写得像千篇一律的 AI，因为大家是能感觉出来的。他们可能不会说这是 AI 写的，也许不会专门回复你点破，但他们就是会没那么信任你。他们会直接划走，不会回复你。甚至可能压根不信你说的话，因为他们会觉得：这是 AI 写的。

[3:52] Greg

So, this repo isn't really about making writing nicer. It's It's really about making your communication more believable. So, if you want to try this one, you can install it You can install it as a skill. That's what I would do. Uh you know, you just type in NPX skills add and then you put the GitHub link and I'll include uh the link to the repo in the show notes in the description and actually all of them, so you can just go and do that. Then you can use it inside your agent workflow, wherever you have a draft of it of something, right? So, the practical way I would run it would be just to, you know, write a rough draft first. I wouldn't ask AI to invent the whole thing from scratch. Uh just, you know, write like a an outline and get the real points down, even if they're messy, and then you would ask the skill to remove the AI sanding patterns while keeping your voice. So, that's the workflow to think about. You have a hu- a human set of ideas first and then you have AI cleanup second, which is different than how a lot of people today are creating content. Uh they're just either asking AI to write everything, basically trying to one-shot it, uh and then and then ask saying, "Hey, this doesn't look good. This doesn't look good."

所以说，这个仓库其实不是为了让文字更漂亮，而是为了让你的沟通更可信。如果你想试试这个，可以把它装成一个 skill，我自己就会这么装。你只要输入 npx skills add，后面跟上 GitHub 链接就行。我会把这个仓库的链接放在节目说明，也就是视频描述里——其实所有仓库的链接都会放，你直接照着装就行。之后你就可以在自己的 agent 工作流里用它，只要手头有什么东西的草稿，都能拿来过一遍，对吧？我实际会这样用：先自己写一版粗糙的草稿。我不会让 AI 从零开始把整篇都编出来，而是先写个大纲，把真正要讲的要点记下来，哪怕乱一点也没关系，然后再让这个 skill 去掉那些 AI 味的套路，同时保留你自己的语气。这就是值得琢磨的工作流：先有一套人的想法，再让 AI 做清理。这跟现在很多人做内容的方式不一样——他们要么让 AI 把所有东西都写了，基本上就是指望一次生成到位（one-shot），然后再一遍遍地说：「哎，这个看着不行。这个看着也不行。」

[5:12] Greg

Um but this sort of like hybrid approach with the skill here, super interesting. Um so, there you go. That's number one. So, the second repo is the is the CRM by trycompai. It's an open-source CRM built for AI agents and that sounds a little technical and maybe boring. I'm going to try to break it down as simple as possible so you can understand how to actually get value out of this if you're building a business. A normal CRM is a place where you store your relationships and your customers and your leads. You know, it's basically a database with names, emails, and notes. Sometimes deal stages if people are using it for sales. The problem is almost all of them depend on you doing the work. So, you have to remember to update the person, you know, remember what happened last time, what they said, who went cold, um and just basically keep your CRM clean. Um and because of that, uh if you're anything like me, you you you've had a CRM and it just turned into these graveyards that you just start to stop updating.

但像这样借助 skill 的混合做法，就特别有意思。好，这就是第一个。第二个仓库是 trycompai 做的 CRM。这是一个专为 AI agent 打造的开源 CRM，听起来有点技术，可能还有点无聊。我会尽量讲得简单点，让你明白如果你在做生意，怎么才能真正从中拿到价值。普通的 CRM 就是一个存放你的人脉关系、客户和销售线索的地方。说白了，它基本上就是个数据库，里面有姓名、邮箱和备注；如果拿来做销售，有时还会有交易阶段。问题是，几乎所有 CRM 都得靠你自己动手。你得记得去更新某个人的情况，记得上次发生了什么、对方说了什么、谁已经没了下文，基本上就是得把 CRM 维护得干干净净。也正因为这样，如果你跟我差不多，你肯定也用过某个 CRM，结果它慢慢变成了一座座坟场，你也渐渐不再更新了。

[6:18] Greg

Um it starts off really exciting and then a few weeks the data goes stale and the notes get messy and you just stop updating it and the people who could actually make you money end up sitting in a tool that nobody wants to open. So, what's interesting about this particular repo is that it treats your CRM as a workspace for an agent instead of a filing cabinet for you. So, the agent can look at records and research people and enrich company info. It could do uh schedule follow-ups and then it could keep the notes current. All the stuff that basically you didn't want, you know, to do. It's it's going to it could do it for you and do those tasks. The whole idea of this of this, you know, AI CRM is that it shouldn't just store information, it should maintain the relationship graph for you. Um you know, it's a big deal if you're a founder um because most businesses actually don't fail from a lack of opportunities. Um a lot of times they fail uh because the opportunities are just scattered everywhere. You get this like warm intro maybe like 6 months ago or you said a you know, you get like uh feedback from a customer who says check back later and you forget about it. Um and there's real money sitting inside those in those emails. So, the use case here isn't to replace uh Salesforce.

一开始你还挺兴奋的，过了几周数据就过时了，备注也乱七八糟，你干脆就不更新了。那些真正能帮你赚钱的人，最后就躺在一个谁都不想打开的工具里。所以这个仓库有意思的地方在于，它把你的 CRM 当成 agent 的工作台，而不是给你自己用的文件柜。agent 可以翻看记录、调研联系人、补全公司信息，还可以安排跟进，并让备注始终保持最新。基本上所有你不想干的活儿，它都能替你干掉。这个 AI CRM 的整个理念就是：它不应该只是存储信息，而应该替你维护整张关系图谱。如果你是创始人，这件事就很关键，因为大多数公司其实并不是因为缺机会而失败的，很多时候失败是因为机会散落得到处都是。比如大概半年前有人给你做过一次熟人引荐，或者有个客户给你反馈说「过段时间再联系」，然后你就把这事忘了。而那些邮件里其实躺着真金白银。所以这里的用法并不是要取代 Salesforce。

[7:44] Greg

Not suggesting that. It's like a smaller useful use case. So, the way I would do it is um say I I'm looking for sponsorships for my podcast. You know, I can run my sponsorship pipeline through it. You know, tracking agency leads. I can manage investor updates or if I have investors or I can follow up with everyone who's joined a waitlist to a something that I've vibe coded or I can have it handle customer success uh just fully for a small SAS. So, if I were testing this, I would actually start with one small workflow. Uh I'd make a pipeline called something like warm leads I can't afford to forget. And then I'd drop in every person who's shown any buying intent and then replied to any email or booked a call or asked about pricing or said they'd circle back, you know, in 45 days um or after Christmas or something. And then I'd let the agent do the hard work to keep that alive. In the past, you'd actually have, you know, you'd hire someone to do this.

我可没这么建议。它更像是一个小而实用的用法。我会这么用：比如我在给我的播客找赞助，就可以把整个赞助商管线（pipeline）放到它上面来跑，跟踪 agency 线索。我也可以用它管理投资人更新——如果我有投资人的话；或者跟进每一个加入了某个我 vibe coding 做出来的产品候补名单的人；又或者让它把一个小型 SaaS 的客户成功工作整个包下来。所以如果要测试这个东西，我其实会先从一个小工作流开始。我会建一条管线，起个类似「绝对不能忘的意向客户」这样的名字。然后把每一个表现出任何购买意向的人都丢进去——回过邮件的、约过电话的、问过价格的，或者说过 45 天后、圣诞节后之类再回头联系的。然后让 agent 去干那些费劲的活，让这些线索一直保持热度。放在以前，你其实得专门雇个人来干这个。

[8:49] Greg

Now you have this agent that's doing it for you. Who needs a follow-up? Who should I say, you know, what should I say, you know, what's changed at their company since we last talked? That's That's why this is really cool. Okay, so how do you install uh this? Well, you're going to need bun and docker installed. So, the basic setup is you're going to want to you know, get clone the repo. Again, I'll have that repo in the in the show notes. You're going to want to do CD CRM uh CP uh env uh dot example.env. You're going to want to do bun install, docker compose up -d, bun run db deploy, bun run db seed, and bun run dev. And I'll include um I'll I'll I'll maybe I'll get some notes up and if you go in the description, I could uh, have that, and I can also show it up, you know, on the screen here, so you can see it. Uh, then what it does is it runs locally. So, you know, usually it runs locally on localhost:3000 with the API on localhost:3001.

而现在有个 agent 替你干这些：谁需要跟进？我该跟谁说——或者说，我该说些什么？自从上次聊过之后，他们公司有什么变化？这就是它很酷的原因。好，那怎么安装呢？首先你需要装好 bun 和 docker。基本的安装步骤是：先 git clone 这个仓库——同样，仓库链接我会放在节目说明里。然后执行 cd crm 和 cp .env.example .env，再执行 bun install、docker compose up -d、bun run db:deploy、bun run db:seed，最后 bun run dev。我会把……我可能会整理一份笔记，你去视频描述里就能看到，我也可以直接把它显示在屏幕上，方便你看。装好之后它就在本地运行，通常跑在 localhost:3000，API 则在 localhost:3001。

[9:53] Greg

Um, so there's environment variables to set, you know, especially for auth, so you want a Google or Microsoft auth client if you want sign-in and email or calendar integrations, which I assume you're going to want. So, this isn't like a one-click, everything is set up beautiful thing, but it's real, it's a product-shaped repo, and I think it's really interesting. Um, I think like the bigger opportunity of something like this is that everyone has a CRM problem, and they just never really call it that, right? They They say, "I forgot to follow up." or "I forgot who to email." Uh, or our leads are a mess. Um, or they have this feeling like, "Oh my god, I'm leaving money on the table." I've certainly felt that. And that's a And that's all the same pain, and this repo just like points at the fix. And like the big idea is how do you create relationship systems that actually do work with you? Agents actually do it. I think it's interesting, and that's why I put it on the list.

另外还得配置一些环境变量，尤其是鉴权（auth）相关的。如果你想要登录功能，以及邮箱或日历集成——我估计你肯定会想要——就需要一个 Google 或 Microsoft 的 auth 客户端。所以它不是那种一键安装、样样都给你配得漂漂亮亮的东西，但它是实打实的，是个有产品样子的仓库，我觉得非常有意思。我觉得这类东西背后更大的机会在于：每个人都有 CRM 问题，只是从来没人真这么叫它，对吧？他们会说「我忘了跟进」，或者「我忘了该给谁发邮件」，或者「我们的线索一团糟」，又或者总有一种感觉：「天哪，我在白白错过赚钱的机会。」我自己肯定有过这种感觉。这些其实都是同一种痛点，而这个仓库正好指向了解法。它背后的大想法是：怎么打造真正能跟你一起干活的关系系统？让 agent 真正把活干了。我觉得这很有意思，所以才把它放进了这份清单。

[10:53] Greg

So, the third repo is by browser use, and it's called video use, and it's one of my favorites because you're going to instantly get it. Um, it lets you edit videos with coding agents. So, you drop your raw footage in a folder, and you can ask an agent like Cloud Code or CodeX to edit it. Uh, it'll remove filler words, it'll cut dead space, you know, it could add subtitles and color grade and build overlays. It could render the video, and it can actually check the output. There already exist a bunch of AI video video editing softwares out there, um, but I think what's really cool about this specific repo is that the editing workflow here becomes something that your agent can understand and repeat. So, think about this. Most creators are any are people who create content have the same handful of edits over and over. They have the same like subtitle style. I mean, even if you look at my YouTube channel, you can see it's like a very similar style every single video.

第三个仓库来自 Browser Use 团队，叫 Video Use，是我最喜欢的仓库之一，因为你一看就懂。它能让你用 coding agent 剪视频。你把原始素材丢进一个文件夹，然后让 Claude Code 或 Codex 这样的 agent 去剪。它会删掉嗯、呃这类口水词，剪掉没内容的空档，还能加字幕、调色、做叠加画面。它能把视频渲染出来，而且还会真的去检查成片。市面上已经有一大堆 AI 视频剪辑软件了，但我觉得这个仓库真正酷的地方在于：剪辑流程在这里变成了你的 agent 能理解、能重复执行的东西。你想想，大多数创作者，或者说任何做内容的人，翻来覆去用的都是那么几种剪法，字幕风格也都是同一套。就拿我的 YouTube 频道来说，你会发现每一期视频的风格都非常相似。

[12:00] Greg

It's at the same certain pace, a certain way of cutting pauses and and opening clips and turning a long conversation into shorts and mids. So, normally all that actually lives in a person's brain and muscle memory. But with a repo like this, it becomes like way more explicit and the agent can actually read the transcript, look at the timestamps, propose an editing strategy, render it, and check the cut points. I mean, how cool is that? Because if you're a creator or founder, content is going to be probably one of your biggest bottlenecks. And you know this, you can have a great product and you're still going to lose if no one's going to see it. So, uh I think this repo is like a way to shrink the distance between raw material and the published content. Um if I were using this, I wouldn't try to automate my whole YouTube channel on day one. I think that's way too much and I think when people use a repo like this on day one and then they start and they don't get the results, it's because they just asked for too much.

节奏是固定的，剪停顿的方式、开场片段的处理、把一段长对话拆成短视频和中视频的方法，也都是固定的。这些东西通常其实只存在某个人的脑子和肌肉记忆里。但有了这样一个仓库，一切就明确多了：agent 真的能读逐字稿、看时间戳、提出剪辑方案、渲染出片，再检查每个剪辑点。这多酷啊？因为如果你是创作者或创始人，内容很可能是你最大的瓶颈之一。你心里也清楚，产品再好，没人看到照样会输。所以我觉得这个仓库就是在缩短从原始素材到发布成品之间的距离。如果换我来用，我不会第一天就想把整个 YouTube 频道全自动化，那步子迈得太大了。我觉得很多人第一天就用上这种仓库，一开跑却没拿到想要的结果，原因就是一上来要求太多了。

[13:11] Greg

I would start with one small repeatable format. So, maybe it's taking a founder's loom and turning it into a 60-second launch video or taking a podcast recording and just being like, I need three banger clips from it. Um that is good enough to start. And then once you have one format, then you can turn it into a system. And then once it's a system, then you can run it yourself or sell it to other people. Startup idea right there, right? Um every niche needs content. And pretty much nobody pretty much no one enjoys editing it. Uh you know, real estate agents need listing videos. SaaS vendors need product uh videos. Uh coaches need clips. And agencies need different ad variations. Um all these, you know, types of companies have different workflow where they're creating content. And it's fun to create content a lot of time, but the actual editing is really hard. Okay, so if you want to try this, how can you try it? Uh it the setup is really agent-friendly.

我会从一个小的、可重复的格式起步。比如把创始人录的一段 Loom 视频做成 60 秒的产品发布视频，或者拿一段播客录音，直接说：我要从里面剪出三条爆款切片。这就足够起步了。等你有了一种格式，就可以把它做成一套系统。一旦成了系统，你既可以自己跑，也可以卖给别人。这不就是一个创业点子吗？每个细分领域都需要内容，而几乎没有人喜欢剪辑。你看，房产经纪需要房源视频，SaaS 厂商需要产品视频，教练需要切片，代理公司需要各种不同版本的广告。这些不同类型的公司做内容的工作流各不相同。很多时候做内容是挺好玩的，但真正的剪辑环节是真难。好，那如果你想试试，要怎么上手呢？它的安装配置对 agent 非常友好。

[14:23] Greg

So all you have to do is paste the setup prompt into Claude code or Codex or whatever, you know, which shell access actually. So you just say setup the GitHub, read the install.md first, install the repo, wire up FFmpeg, and then register the skill with the agent I'm using. And then ask me for the 11 Labs API key when needed. Um you know, could you install this manually? Yes, you can do get clone and ln sfn and you know, all that stuff, but you know, for a lot for 95% of people listening to this, you know, might as well just have the agent install it for you. Um that's that's a an interesting one. I feel like try it, see if it works for you, and uh let me let me know what you think. So the fourth repo is by Nvidia and it's called Skill Specter. Um this this is a no-brainer one to install because people are starting to install skills and MCP servers and plugins and agent tools straight from GitHub, which is exciting because it gives your AI new abilities, but it creates a problem because a skill isn't just a block of text, right? It can include instructions and scripts and dependencies, tool access. Uh and basically behaviors that change how your agent works. And in some cases, it can actually touch files, call tools, and connect to services.

你只需要把 setup prompt 粘贴进 Claude Code、Codex，或者随便哪个真正有 shell 权限的 agent 里。你就这么说：帮我把这个 GitHub 仓库装好，先读 install.md，安装仓库，配好 FFmpeg，再把这个 skill 注册到我正在用的 agent 上，需要的时候再问我要 ElevenLabs 的 API key。那能不能手动装？当然能，你可以 git clone、ln -sfn 之类的一通操作，但对 95% 听这期节目的人来说，还不如直接让 agent 替你装。这个仓库挺有意思的。我建议你试试，看看它适不适合你，也告诉我你怎么看。第四个仓库出自 NVIDIA，叫 SkillSpector。这个是想都不用想就该装的，因为大家已经开始直接从 GitHub 安装 skill、MCP server、plugin 和各种 agent 工具了。这件事令人兴奋，因为它能给你的 AI 带来新能力，但也带来了一个问题：skill 可不只是一段文字，对吧？它可能包含指令、脚本、依赖和工具权限，说白了就是各种会改变你 agent 工作方式的行为。在某些情况下，它真的能动你的文件、调用工具、连接外部服务。

[15:58] Greg

So, before you install random capabilities from GitHub, you should probably ask whether it's safe, and that's what Skill Specter is for. It can scan AI agent skills for security issues, uh things like prompt injection, data exfiltration, supply chain risk, hidden instructions, basically any malicious patterns, um and MCP-related risks, uh Skill Specter has your back. So, the way I think think about it is before you hand your AI a new tool, scan the tool, right? Um and this matters uh a lot more now because AI workflows are getting more modular. So, most people use one chatbot today, but builders are already assembling their own AI work environments like a coding agent, a research skill, a design skill, maybe it's a a browser tool, a video workflow. And at some point, your setup starts to look like a little operating system for your work. And once that happens, security stops being an enterprise-only problem and becomes a normal like builder or founder problem. So, I think it's a really important one. No affiliation with uh Nvidia.

所以在你从 GitHub 上随手安装各种来路不明的能力之前，最好先问一句：这东西安全吗？SkillSpector 就是干这个的。它能扫描 AI agent 的 skill 有没有安全问题，比如 prompt 注入、数据外泄、供应链风险、隐藏指令，基本上任何恶意模式，还有跟 MCP 相关的风险，SkillSpector 都替你把关。我的思路是：在把新工具交给你的 AI 之前，先扫一遍这个工具，对吧？而且这件事现在重要得多，因为 AI 工作流正变得越来越模块化。如今大多数人只用一个聊天机器人，但 builder 们已经在拼装自己的 AI 工作环境了，比如一个 coding agent、一个调研 skill、一个设计 skill，可能还有一个浏览器工具、一套视频工作流。到了某个时候，你这一套配置看起来就像一个专为你的工作服务的小型操作系统。一旦到了这一步，安全就不再只是企业才需要操心的事，而会变成普通 builder 或创始人都得面对的问题。所以我觉得这个仓库真的很重要。声明一下，我跟 NVIDIA 没有任何利益关系。

[17:17] Greg

But uh if you want to try it, uh you know, you can use UV. The quick install is UV tool install get plus, and then you just post the GitHub uh uh link. Um and then just scan a skill directory with skill spector scan {dot} {slash} my {dash} skill, or you could just scan a GitHub repo directly with skill spector scan and then you with the link. Um if you want a faster stat static scan that doesn't send file contents to an LM provider, if you add {dash} {dash} no {dash} LM, you know, it matters if you're scanning like sensitive files or or or private files in general. Uh there's also a docker option if you'd rather not install Python locally, which I think some people, you know, would want to know that. I've been thinking about this a bunch. Like if teams are going to install AI skills and MCP servers, someone's going to need to help them decide what's safe. You know, that could be a trusted marketplace, a security scanner, uh an install gate for companies, or a feature inside every agent platform. I like that Nvidia is attached to this one. That's why I included it uh because it gives the whole category credibility.

如果你想试试，可以用 uv。快速安装命令是 uv tool install git+，后面直接贴上 GitHub 链接就行。然后用 skillspector scan ./my-skill 扫描一个 skill 目录，或者直接用 skillspector scan 加上链接，扫描一个 GitHub 仓库。如果你想要更快的静态扫描、不把文件内容发给 LLM 服务商，就加上 --no-llm。要是你扫的是敏感文件，或者总体来说是私密文件，这一点就很重要。如果你不想在本地装 Python，还有 Docker 方案可选，我想有些人会想知道这一点。这件事我琢磨了挺久。如果团队要安装 AI skill 和 MCP server，总得有人帮他们判断什么是安全的。这可以是一个可信的应用市场、一个安全扫描器、一道面向企业的安装关卡，或者是每个 agent 平台里都内置的一项功能。我喜欢这个项目有 NVIDIA 背书，这也是我把它收进来的原因，因为它给整个品类带来了可信度。

[18:35] Greg

And it also is just admitting that, hey, like a lot of these skills, the plugins, things like that, could be malicious, and you got to be careful. So, I had to put that in there. Had to keep it real with you all. And uh check it out. So, the fifth repo is called phone harness, and it lets AI lets an AI agent actually control your phone. Um most AI most AI agent tools are built around browsers and terminals. Um also desktop apps, obviously makes sense because a lot of the work happens there. But, think about it. a huge amount of work happens on phones now, and banking apps, and messaging, and social, and food delivery, all that sort of stuff. So, there's a ton of workflows that are hard to automate because the only real interface is a phone screen. So, Phone Harness connects an agent like Codex or Cloud Code to a real iPhone or Android. Super cool, right? For iPhone, it uses Mac iPhone mirroring, and for Android, it uses ADB. So, you don't have to like jailbreak it, which, you know, I was like, "Oh my god, I don't want to jailbreak anything." You don't have to jailbreak it. You don't, you know, no Xcode. Um you don't really need like you don't need like a special app on the phone, either. Basically, the agent sees what's on the screen, and it could tap, and it could type, and it could scroll, and it could open app, and open apps,

而且它也等于坦白承认：嘿，很多这类 skill、plugin 之类的东西可能是恶意的，你得小心。所以我必须把它放进来，得跟大家说实话。去看看吧。第五个仓库叫 Phone Harness，它能让 AI agent 真正操控你的手机。大多数 AI agent 工具都是围绕浏览器和终端打造的，还有桌面应用，这显然说得通，因为大量工作都在那些地方完成。但你想想，现在有海量的事情是在手机上完成的：银行 App、聊天消息、社交、外卖，诸如此类。所以有一大堆工作流很难自动化，因为唯一真正能操作的界面就是手机屏幕。Phone Harness 能把 Codex 或 Claude Code 这样的 agent 连到一台真实的 iPhone 或 Android 手机上。超酷，对吧？iPhone 用的是 Mac 上的 iPhone 镜像功能，Android 用的是 ADB。所以你不需要越狱，我当时心想：天哪，我可不想越狱任何东西。你不用越狱，也不用 Xcode，手机上也不需要装什么特别的 App。基本上就是 agent 能看到屏幕上的内容，然后可以点击、打字、滚动、打开 App，打开各种 App，

[20:06] Greg

and then it verifies what happens. I know there's someone, you know, thinking like, "Why? Why does that matter?" Well, if you're building a mobile app, think about the QA, right? You can have an agent test onboarding, sign up, checkout on an actual real device. If you're building any sort of automations, it could open up mobile apps that have no API. And if you're a creator or even an operator, it could eventually handle repetitive mobile workflows across things like TikTok and Instagram. So, if you want to try this, the first thing uh to do uh is test a mobile sign up flow. Super simple. Open up the app, create an account, tap through onboarding, try checkout, take screenshots, and tell me where it gets confusing or breaks. That's super valuable. Every mobile team should be constantly doing that, but most don't because it's tedious, or they don't want to hire that person to go and do it. Or they honestly like make mistakes. So, this is going to uh help you there. If you want to try this one, uh the repo gives you a setup prompt to paste into Codex or Cloud Code.

然后核实操作之后发生了什么。我知道肯定有人在想：为什么？这有什么意义？好，如果你在做移动 App，想想 QA，对吧？你可以让 agent 在一台真实设备上测试新手引导、注册和结账流程。如果你在做任何自动化，它能打开那些没有 API 的手机 App。如果你是创作者，甚至是做运营的，它最终还能处理 TikTok、Instagram 这类平台上重复性的手机端工作流。所以如果你想试试，第一件事就是测一遍手机端的注册流程。超级简单：打开 App，创建账号，把新手引导一路点完，试一下结账，截图，然后告诉我哪里让人困惑、哪里会出故障。这非常有价值。每个移动端团队都应该经常做这件事，但大多数团队都没做，因为太繁琐了，或者他们不想专门雇个人去干这个，又或者说实话，他们自己也会出错。所以这个工具能在这方面帮到你。如果你想试这个，仓库里给了一段 setup prompt，粘贴进 Codex 或 Claude Code 就行。

[21:18] Greg

Uh I'll include it uh I'll include it. Basically, set up phone harness for me, clone the GitHub, read the install.md, install uh install it so phone harness is on my path, register it as an agent skill, and then read onboarding.md, and walk me through the setup. Uh for iPhone, you do need Mac OS Sequoia or later with iPhone mirroring plus accessibility and screen recording permissions granted to your terminal. Uh for Android from Android friends, turn on developer options and connect through ADB over USB or it actually works wireless, too. Then, run phone harness --doctor to check the setup. Um it's early. This one is early, and it has its limits. Um some phone flows won't work perfectly. Uh for example, like lock phones need you, and Face ID and camera flows are pretty tricky. But, the direction is obvious, right? Agents are moving from answering questions, they're operating tools, and phones is one of the biggest tool surfaces in the world. And, you know, if you figure out some of these workflows, um there's startup ideas right there, right? If you can uh productize mobile QA, would would someone be willing to pay $100 a month, $500 a month for that? How many clients do you need to get to $10,000 a month?

我会把它附上。大意就是：帮我装好 Phone Harness，克隆这个 GitHub 仓库，读 install.md，安装好并确保 phone-harness 在我的 PATH 里，把它注册成一个 agent skill，然后读 onboarding.md，一步步带我完成配置。iPhone 这边，你得有 macOS Sequoia 或更新版本，开启 iPhone 镜像，还要给你的终端授予辅助功能和屏幕录制权限。用 Android 的朋友们，打开开发者选项，通过 USB 用 ADB 连接，其实无线连接也行。然后运行 phone-harness --doctor 检查配置。这个项目还很早期，有它的局限，有些手机上的流程没法完美跑通。比如锁屏的手机还得你亲自来，Face ID 和相机相关的流程也挺棘手。但方向是显而易见的，对吧？agent 正在从回答问题转向操作工具，而手机是全世界最大的工具界面之一。而且如果你把其中一些工作流摸透了，创业点子不就在这儿吗？如果你能把移动端 QA 做成产品，会有人愿意每月为此付 100 美元、500 美元吗？你需要多少客户才能做到每月 1 万美元？

[22:44] Greg

Um there's That's in general what's interesting about a lot of these uh GitHub repos is, you know, you can you can install them. Step two is have a small workflow, uh get it to work, add value, and step three is then, should I be productizing this for other people or just continuing continuing using this in my own workflows to be more productive to make money and create value. So there you have it folks, five GitHub GitHub repos I think are really interesting. Obviously I don't want to be a full-time open source analyst but because GitHub is one of the few places you can still find genuinely useful tools before they go mainstream you got to be spending time there. I think it's totally worth it. Um and as you install some of these repos um even if you're not technical, you learn you learn things, right? You you start you start learning what works for you, uh what's what's this repo missing, uh how how can you make this work for your company, how can you sell this to other people?

总的来说，这类 GitHub 仓库很多都有个有意思的地方：你可以先把它们装上；第二步，搭一个小工作流，把它跑通、创造价值；第三步再想想，我是该把它做成产品给别人用，还是继续在自己的工作流里用它，提高效率、赚钱、创造价值。好了朋友们，这就是我觉得非常有意思的五个 GitHub 仓库。当然，我可不想当个全职的开源分析师，但 GitHub 是少数几个还能在好工具火起来之前就发现它们的地方之一，所以你得在上面花点时间。我觉得完全值得。而且在你安装这些仓库的过程中，就算你不懂技术，也能学到东西，对吧？你会开始摸清什么适合你、这个仓库还缺什么、怎么让它在你的公司里用起来、怎么把它卖给别人。

[23:50] Greg

Um I just think it's it's a really interesting thing to do do to just do like, you know, monthly uh finding new GitHub repos, installing them uh you know, you'll be able to build faster and uh it'll just connect the dots in your head in this agentic era that I think it's worth it. So um I don't know if this has been fun for you, interesting for you, valuable for you. If you listen to this podcast, you know I live to serve and all I care about is just increasing your probability of success in this era with ideas, with tools, with frameworks. So please let me know in the comments what you thought of this. Should I be doing this every month, every 2 weeks, every 6 weeks, never again? All good. I appreciate you so much um and uh I'm rooting for you. So I'll see you next time and uh have a creative day.

我就是觉得这件事特别值得做：每个月找找新的 GitHub 仓库，装上试试。这样你能更快地把东西做出来，而且在这个 agentic 时代，它会帮你在脑子里把各个点串起来，我觉得很值。所以，不知道这期对你来说好不好玩、有没有意思、有没有价值。如果你常听这个播客，就知道我活着就是为了服务大家，我唯一在乎的，就是用点子、工具和框架，提高你在这个时代成功的概率。所以请在评论区告诉我你觉得这期怎么样。我是该每个月做一期、每两周一期、每六周一期，还是再也别做了？怎么都行。真的非常感谢你，我为你加油。下期见，祝你度过充满创造力的一天。