Copy any app, Claude does SEO, and the top 10 repos
A two-host weekly walk through the ten most-starred GitHub repos for AI builders, plus three bonus picks, with a running argument about which open-source tools can ever become businesses.
Posted
2 days ago
Duration
Format
Listicle
educational
Views
21.1K
407 likes
57 · 43
Big Idea
The argument in one line.
The best agent skills encode a point of view and an order of operations the model already has the raw ability for, but because they are trivially copyable and age with each model release, their value lies in curation, freshness and the community around them.
Who This Is For
Read if. Skip if.
READ IF YOU ARE…
A builder working in Claude Code, Codex or Cursor who wants a short, opinionated list of skills and repos worth installing this week.
Someone deciding between free crawl-based SEO tooling and paid keyword data for a new site.
A developer juggling several coding agents across multiple machines who is looking for one interface to manage them.
A founder or investor trying to judge whether a popular open-source AI tool can turn into a defensible company.
SKIP IF…
You want hands-on setup walkthroughs; the repos are shown as summary cards and discussed, not installed step by step.
You already track GitHub trending daily; most of the value here is commentary, not discovery.
TL;DR
The full version, fast.
Ten trending GitHub repos plus three bonus picks show where agent tooling is heading: packaged skills that encode an expert's order of operations. Cloudflare's security-audit skill earns praise for flagging fewer, more relevant issues, Claude SEO swaps paid DataForSEO data for a free site crawl plus Google APIs, and Superpowers applies Cialdini's persuasion principles to make agents comply. The hosts test each pick against two filters. First, check the last commit, because a skill tuned for models from six months ago may underperform on today's. Second, ask what stops someone copying it for free; for most of these tools the honest answer is distribution, community and services, not the code.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
An open-source launcher that hosts self-built clones of business apps on a $5/month Cloudflare Workers plan and shares them with a team, compared with pricier hosts like Fly.io.
02:16 – 03:43
02 · Bonus: tstack, a skeptic's skill pack
A developer who mocked skills publishes his own nine. The co-host describes himself as a skill realist: install fewer, read them, customize them.
03:43 – 05:43
03 · Bonus: iPhone apps on Linux
An Omarchy community tool builds and signs iOS apps without a Mac. Why Apple requires Macs, legally and through Xcode, and why this may be a long cat-and-mouse fight.
05:43 – 07:00
04 · #1 Cloudflare security-audit skill
Audits repos before you install them or audits your own agent-written code. Praised for flagging only what matters instead of a nitpick list.
07:00 – 09:31
05 · #2 Claude SEO vs OpenSEO
A full SEO audit skill that runs on free crawl data and Google APIs, with paid data optional. Where free data runs out, and the undercutting chain from Ahrefs to OpenSEO to Claude SEO.
09:31 – 12:40
06 · #3 Impeccable and the moat question
A design skill that learns your taste first and takes click-and-draw feedback. At 96K stars it looks like a free Figma killer, which raises the question of what any of these projects can charge for.
12:40 – 14:35
07 · Sponsor: TripleTen
AI systems engineering and machine learning programs taught by working engineers, framed against a reported fourfold rise in job cuts attributed to AI.
14:35 – 17:10
08 · #4 Superpowers
A coding skill pack that applies Cialdini's persuasion principles to agents, citing a Wharton study where persuasion more than doubled model compliance.
17:10 – 18:40
09 · #5 TradingAgents
Multi-agent stock debate on simulated money. Most viral 'I gave my agent $1,000' posts are fake money, and real trading desks are far more advanced.
18:40 – 21:43
10 · #6 Karpathy-inspired rules file
A 226K-star rules file whose last commit is six months old, which prompts a warning about stale skills. Then a detour on why agents abandon external task boards.
21:43 – 22:37
11 · Sponsor: Zapier
An agent that reviews routines and moves repeatable workflows into cheaper Zaps, plus Zapier's free-forever plan.
22:37 – 24:35
12 · #7 Taste skill
An art-director skill for coding agents. One host runs several design skills against the same site for ideas; the other prefers screenshots over formulaic style presets.
24:35 – 26:19
13 · #8 Interface design skills
Fourteen skills that teach agents interface design and accessibility, filling the gap where agents ship screens with too many buttons and too much text.
26:19 – 30:46
14 · #9 T3 Code
A meta harness over Claude Code, Codex and Cursor with synced projects across machines and phone, settled threads, thread references and scheduled prompts for when limits reset.
30:46 – 31:44
15 · #10 Anthropic knowledge-work plugins
Anthropic's plugin set that turns Claude into a specialist for sales, productivity and other job functions.
31:44 – 34:24
16 · What the hosts built
A context-layer app built in five to ten hours for about $100 a month, and a skill that edits each recorded episode from a Riverside link in the show's house style.
Atomic Insights
Lines worth screenshotting.
Most open-source AI apps die on the laptop they were installed on; cheap hosting like a $5/month Cloudflare Workers plan is what makes them shareable with a team.
A skill realist installs few skills, reads every one, and customizes them instead of piling up dozens of untouched defaults.
Security-audit tools that return long nitpicky lists cost you twice, because every fix your agent makes adds code that can break later.
Free crawl data is enough to start on SEO, but only paid keyword data tells you whether anyone actually searches for the term you want to rank for.
A skill often teaches a model nothing new; its value is the order in which to run the steps and the point of view baked in.
OpenSEO is doing to Ahrefs what Claude SEO is now doing to OpenSEO, so each layer of cheaper tooling gets undercut by the next.
When any team can build an internal version of a design skill in an afternoon, the open-source author's likeliest business is selling services to the users the tool attracted.
Persuasion principles from Cialdini's Influence, such as commitment and social proof, measurably raise how often language models comply with instructions.
Agent trading results posted online are almost always simulated money, and an individual's open-source bot is competing against professional trading desks.
Check a skill's last commit before installing it; prompts written for models from six months ago may not suit the current generation.
Agents drift away from external task boards; to-do tracking works best when it is built into the agent harness itself.
Design skills are most useful as idea generators: run three of them against the same page, keep the parts you like, and discard the rest.
Coding agents still produce interfaces with too many buttons and too much text, so a dedicated interface-design and accessibility skill fills a real gap.
A meta harness sits above Claude Code, Codex and Cursor, letting you run all of them from one interface and keep the same chats synced across machines and your phone.
GitHub stars lag real usage, because the busy people relying on a tool every day often never star it.
An internal context-layer app that ingests email, Slack, calendar, docs and call transcripts took five to ten hours to build and runs for about $100 a month in compute.
Takeaway
Judge every skill by freshness, focus and moat
WHAT TO LEARN
The most useful agent tools package an expert's order of operations, but you should check how current they are, keep only the ones you read, and expect easy copies.
01Bonus: Every App
Self-hosted open-source apps only pay off once they run somewhere always on and shareable, not on a laptop that sleeps.
Compare hosting cost per app: one n8n instance on a pricier host can run $20 to $30 a month, while a Workers plan starts near $5.
02Bonus: tstack
Treat published skill packs as a starting point; keep the few that help and customize the rest before relying on them.
If you have not read a skill you installed, you do not know what it is telling your agent to do.
03Bonus: iPhone apps on Linux
Apple ties iOS signing and App Store Connect uploads to Xcode on a Mac, both technically and in its terms of service.
Workarounds that reverse-engineer a platform owner's toolchain carry ongoing breakage risk, so plan for a maintenance fight.
04#1 Cloudflare security-audit skill
Run a security audit on third-party repos before installing them and on agent-written code before shipping it.
Prefer auditors that rank findings, since fixing every nitpick adds code and complexity that create new failure points.
05#2 Claude SEO vs OpenSEO
Start SEO with free crawl data and Google APIs; it is far better than doing nothing.
Add paid keyword data once you need to know search volume, because wanting to rank for a term does not mean anyone searches it.
A skill's value is often its sequence and point of view rather than any capability the model lacks.
06#3 Impeccable and the moat question
Design tools that interview you about your taste first produce work aimed at your goals rather than the model's defaults.
Direct on-screen feedback, clicking or drawing on what to change, speeds up iteration with an agent.
When a team can build a simpler internal version in-house, open-source tools usually monetize through services, community or distribution.
08#4 Superpowers
Write agent instructions as persuasion: ask the agent to commit to a step before it acts, then hold it to that commitment.
Clear communication habits that work on people, stating intent and confirming understanding, carry over to prompting.
09#5 TradingAgents
Assume viral AI trading results are simulated unless real money is explicitly shown.
An individual's open-source trading agent competes against professional desks with far more sophisticated systems.
10#6 Karpathy-inspired rules file
Check a skill's last commit and rewrite stale instructions for the model generation you actually use.
Task tracking sticks when it is built into the agent harness; external boards get abandoned as agents drift.
12#7 Taste skill
Run several design skills against the same page to generate options, then keep what works and discard the rest.
Screenshots of designs you admire can communicate taste better than a formulaic style preset.
13#8 Interface design skills
Agents still overload screens with buttons and text, so review agent-built interfaces from a real user's point of view.
Bake accessibility, such as alt text and page hierarchy, into the build step instead of auditing it afterward.
14#9 T3 Code
A meta harness lets you keep paying for the agents you already use while managing them from one synced interface.
Re-test a tool you abandoned months ago, because fast-moving AI projects can change completely in a quarter.
16What the hosts built
A context layer over email, chat, calendar, documents and calls can be built in hours and gives one queryable view of every relationship.
Teach an editing skill to expect your feedback so revision passes adjust the existing cut instead of rerunning everything.
Glossary
Terms worth knowing.
Agent skill
A packaged set of instructions, and sometimes scripts, that a coding agent such as Claude Code loads to perform a specific task in a consistent, expert-designed way.
DataForSEO
A paid API that sells search-engine data such as keyword volumes, rankings and backlinks per request, used behind many SEO tools.
App Store Connect
Apple's developer portal where signed iPhone apps are uploaded for review and listing in the App Store.
Meta harness
A layer that controls several coding-agent harnesses, such as Claude Code, Codex and Cursor, from one interface instead of replacing any of them.
Omarchy
An opinionated Linux setup whose community builds Linux versions of common Mac-only tools, aimed at an agent-first developer workflow.
AI native services
A business model that pairs a small amount of software with an outcome-based service delivered on top of it, rather than selling the software alone.
Context layer
An internal system that pulls in a company's email, chat, calendar, documents and call transcripts so people and agents can query everything about a person or company in one place.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
17px
metaphoranalogystory
Presented by Zapier, the AI automation company. Before I get into the top 10, Adam, you got to see this one. This is called Every App.
It's from the guy who created OpenSEO. And he now said, we are going to cop or at least open source every app on the planet that's worthwhile. But how do you host it?
So we created this thing called Every App. It basically allows you to easily host your stuff on Cloudflare's workers paid plan, which is like five bucks a month, super cheap, and then make it accessible to everyone on your team. This is his plan.
What do you think? I mean, I think it's awesome that this is clearly the hardest part of all this cool open source stuff is you run it on your laptop and it's cool and you're like, great. Oh, no, I closed my laptop.
I'm driving in my car like the thing's not running. I can't put it somewhere. How do I share this with my friends?
So I love this. I love that it's not just a tool to help you deploy there, but that it's also this like list of here's the tools that are already ready to go. That, I think, makes a big difference because there's so much cool stuff out there.
And for new users, it's hard to install. So I like it. I like that it's on such a cheap platform.
We've seen this on other platforms like Fly .io in the past, but it's very expensive. If you want to spin up your own N8n or something, you could do that. It's open source, but you're going to be paying $20, $30 a month just for that one N8n instance.
let alone if you install three or four things. Before we go to the next one, what are you wearing? What is this?
It's always tea time. I am the Mad Hatter. I'm going to my kid's school later.
We've got some stuff. We like to dress up in the office. We've got an Elizabeth Holmes running around.
There's a Bob Ross. There's a couple other people. I love that you're doing it even for this.
Tomorrow, Halloween, of course. Next. T -Stack from our guy Theo, he has decided, okay, screw it.
I've been making fun of skills. I think he's been making fun of skills a little bit. Then he accepted other people's skills, and now he's saying, you know what?
I'm going to make my skills available. Adam, you showed me a skills file that you have, so even you, who has poo -pooed skills on this show, likes it. Here is his basically collection of skills that he says that he uses.
I think it was like 70 % of the time that he publishes code. Nice. Yeah.
Yeah, it's cool. I actually haven't watched his skills video on this, but I follow a lot of his content. I bet if you go to my YouTube, I've listened to probably 80 % of his channel.
And yeah, he's kind of... I wouldn't say he's against skills, and I'm not against skills. I would say Theo and I, you know, we're buds, we are skill realists, which is you've probably got too many skills, and if you're not reading the skills that you're installing, if you're not customizing them...
then you're not really doing it right. And I guarantee he covers that in his video. He talks about this a lot in other videos.
And he covered like Matt Pocock's. He covered potato skills also. He says the same things there.
It's like, hey, these three seem to be really useful for me. These ones, not so much. This one I'd love to customize.
And so I like that he's pushing some out, but I also like. his measured approach to how people should adopt these skills we're going to see more from him later today but finally before we start my third one to show you is omachi apple dev and the reason i want to show you this is because someone created a way to build iPhone apps on Linux with no Mac.
And we'll talk about how this is done. You and I talked about it before. The reason I'm showing this is not because of this specific skill or this specific tool.
It's because the thing about the Omachi community is they're creating a brand new operating system that's built for agents first on Linux. And they are so rabid at creating all the tools that you need to make it work. So I said to one person, What about Granola?
I need a meeting note taker. And they said, look, this guy created Granola for Linux. All of these things are available.
How does this work, though? Apple doesn't want you building iPhone apps on Linux. Yeah, they really don't.
And actually, I've been at companies where in order to deploy and test iOS apps, you know, everybody runs Linux servers, but somewhere there's a closet and there's 15 Mac studios or something in the closet just to generate with Apple's proprietary fonts or to use Xcode. And that is a requirement. It's a requirement on two levels.
One is it's a legal requirement. It's in the terms of service. If you're building and signing with Apple certificates, you need to be on a Mac.
The second is the one that they've gotten over here, which is Xcode only runs on Mac, and it's the only way to sign these apps. It's the only way to get them pushed up into the App Store, like what's called App Store Connect, which is where developers upload their apps for approval and for listing in the store. I don't know exactly.
I don't know exactly how they got around that. I'm sure Apple tries to prevent you from any next code or from extracting the certificates from it. So I'm guessing that they're going to be kind of fighting a long -term battle here.
But hey, with AI tools, it's pretty easy to keep catching up, right? If you're reverse engineering, if you're doing a lot of scraping, maybe these guys can stay ahead. Even if they lose on this, I want them to do well in general.
Next. Yeah. The number one repo for the week.
is Cloudflare's security audit skill. You know when you're looking at all these things on GitHub and you're wondering, is this safe to install? Well, this will help you decide that.
Or when you're creating software, you're saying, did I screw up? Did my agent leave a big security hole somewhere? This is going to help you find it.
It is from Cloudflare, a team that is known for security, and it's been incredibly popular this week. Look at this chart. Yeah, I've actually tried this one, and I think what I liked about it was that it seemed...
It didn't inundate me with... useless crap. A lot of these that I've tried, they come back with the longest list of such nitpicky and not helpful feedback.
And I can't tell in that list what's actually worth fixing and what's not. And of course, I have an agent. The agent can work through all of it.
But every time you fix one of these bugs, it adds a little bit more complexity. Now you write more code. Now there's more code that can break.
It's harder to build around. So this one seemed, I don't know why, but it seemed to be a little bit more sane about what it thought you should review and what wasn't worth looking at.
All right. Let's go on to number two, which is Claude SEO. This is a full SEO audit inside a Claude code.
At some point, they're going to be told to get rid of the name Claude from their SEO, from their name. It is doing really well this week. I want to compare it for a moment to OpenSEO because we've talked about this a lot before.
OpenSEO really depends on data for SEO to work. We're talking about a few pennies per week to use this. It's not an expensive tool.
It's used by, I think, the professional SEO software companies. But here, this is a direct connection to data for SEO with a beautiful interface to make it usable and to make it so that you can get better SEO decisions. What they've decided to do is allow you to not...
Pay for that, and it'll crawl your site. You'll get free Google APIs for data, public sources, and if you want to pay, you can pay. By the way, Firecrawl is one of the people you can pay for, and I've got the founder of Firecrawl.
He's coming out, so I hope people subscribe for that. It's going well. Do you believe in something like this?
Is free data enough? I think it's fine to start. If you're doing nothing and you want to improve your SEO, then yeah, free data is way better.
And Claude can do all these things. When I first saw this, my number one question, is this a skill or a tool? And it's a skill.
So clearly, Claude can do all of this stuff already if you would just instruct it. But man, there's a technique to it. There's an order that you should run these things.
And there's a point of view. And that's all baked into the skill here. Now, at some point, though, where SEO runs out is you say, well, I'm looking at my site.
I'm generating new content. What you actually need to do is figure out. What are people actually searching for?
What are other sites ranking for? And that's where the paid data comes in. You have to layer that on because that's how you know where there's volume.
Just because you want to rank for some specific keyword doesn't mean people are searching for that keyword. And you need access to data to know whether or not people are searching for it. The interesting thing here is that OpenSEO is doing to Ahrefs what these guys are doing to OpenSEO.
Yeah, it's hilarious. Where OpenSEO might ultimately win is They do have a community of people on Slack that are customers who are then sharing how they're using it.
That is valuable. OpenSEO also has Ben's fame, the fact that he is out there and people talking about him. And now that he's in Y Combinator, I think that's also going to help him.
And I'm looking at what the moats are, how do you protect the business, and maybe it's those kinds of things. All right, now number three, Impeccable. I did a full interview with the founder of Impeccable.
Here it is. this is beautiful what i like about impeccable is they will they will uh sorry the once you install it it will work with you to understand what you think is good design what you're trying to accomplish and then let me go over to their website here so you can see some of their some of their style docs um And then it will create something for you based on your needs, not its own personal taste.
And then after that, Adam, I really like how you can give your agent feedback where you can just. Touch a thing on the screen and say, I don't like it. You can draw on your screen and say, I want this moved over there.
That little bit of feedback is super helpful for working with your agent. And that was something that I hadn't. Here it is right here.
You can see how you're interacting with the agent to give it feedback. It's something that I hadn't pointed out before because I didn't notice it. I mean, as I watch you click around in here, Andrew, I'm just thinking, and this is free?
I know. I can use this for free. He is working on turning this into a business.
He's got over 95 ,000 stars now. Yeah. He's working on turning it into a business.
OpenSEO got funded by Y Combinator. They're turning it into a business. I think it's challenging.
Yeah. I mean, I could see how it's challenging because, you know, this is probably, I don't know if he would describe it this way, but it looks to me like a Figma killer. Right.
Like I used to use Figma. We prototype, we design in it. It's like, oh, you could do all that way better.
And now you're telling me for free inside Claude. And the problem is like. Yeah, there can be an impeccable killer too.
It's so easy to compete. Like internally at GatewayX, we have our own skill that does, it's nowhere near as detailed as this. It's not as design oriented, but it's around mock -ups and flows and you can draw boxes on the screen and give feedback.
And that's just a skill we all have running on our computers. It's all free to run. No one can ever charge us for it, except for Claude.
And so, yeah, like I wonder as these things try to turn into businesses. What is the actual moat there? And a lot of them, I think, honestly, Andrew, they just don't have it.
They may not. It's great open source software. But maybe if I'm the founder here, maybe I sell my services.
It's like, I've got this great platform, you know me, you trust me, and you can have a really wide distribution. If he has millions of people using this, well, he's going to have at least tens of thousands that want to pay him for design work or for help or something custom. So that might be the business model.
We play a lot in what are called AI native services, which is some blend of these two things, a little bit of software, but then also providing an outcome -based service on top. Yeah, you mean you as the venture studio investing in companies. GatewayX is the name of this venture studio.
By the way, I keep linking to your LinkedIn. I hope people are slamming you on LinkedIn. You are not, like, super active on it.
I'm getting messages. Good. Send Adam more messages.
I'll keep doing it. And I always remember, like, Saturday morning, I wake up and I go, oh, wait, I've got to add Adam somewhere prominent for people to link to. Okay, this is a sponsor.
It is called Triple Ten. This is the projects from their students. There literally are hundreds of them.
All I could do is squeeze these on the page. The reason I'm bringing this up is because they've got working engineers teaching you how to build and ship. And you can go to GitHub and you can see what their students have shipped and play around with their stuff to see it.
The idea behind this is, listen, you're now in a market where AI is just taking over a lot of jobs. Look at this. So far this year, over four times, or about four times exactly, the number of job cuts have been blamed on AI as last year.
And last year was big also. So there is clearly an issue here where people are losing their jobs to AI. But there's also an opportunity for people who know how to use this really, really well.
And so what Triple Ten said is we are going to bring in the best teachers, people who have experience and are currently working for and hiring at companies like Microsoft and Google. Also, they are backed by a publicly traded company called Nebius. And so they've got their practicing engineers at that company.
They're putting on their training programs for people like you. This is not easy. It is not meant for like a YouTuber like me to teach you.
It is meant for professionals, working engineers to teach you how to build and more importantly, how to ship products at them. And they've got these two programs that they suggested that we tell people about. I'm just going to put links to them below.
You've got the AI systems engineer for engineers who already ship production code. You're going to go from feature level engineer to production AI architect with five production grade projects on one real system, each defended live before a mentor. I mean, we're talking about very rigorous programs.
If you're just looking to sit back and watch YouTube and this is like an alternative, it's not. Or they also have the AI machine learning program. Again, links and information below.
Contact them and they will help you out. I am really proud of them. Adam, I had this on last week.
I deleted it. I did not like how I talked about it last week because I felt like it just didn't have the heart that I wanted. Anyway, so today we're going to keep it in.
Next, Obra Superpowers. This is a collection of skills to help you code.
And you're going to say, I already anticipated you, Adam. In my head, I go, what's different about this, Andrew? What's different about this?
The philosophy is kind of consistent to something Adam has talked about in the past, testive and development we've seen before, systematic over ad hoc, process over guessing, complexity reduction. So far, all of you all are listening and you're going, I don't see it. So I put this together because I looked at Jesse, Jesse Vincent, the guy behind this, to see what is his philosophy.
This is a guy who read Robert Cialdini's book on persuasion and said, how do I use... influence the book, to influence my agents to do the freaking thing they're supposed to do. And so he will get them to do things like commit to doing it and then know that they will follow through on their commitments.
Social proof. Anyway, it's one of the many things that he has. He has this whole, there's a whole bunch of posts that he has written online.
This is the guy who is creating this skill that has taken off. What do you think of this? It's not often we see something for the first time.
And this is something I'm seeing for the first time. Like, wow, what a unique approach. You know, yeah, we've seen tons of TDD skill packs out there.
Fine. I'm sure they're all slightly different and some are better. Like, oh, I get this one.
I see why this is different. First of all, I love that book. Like, everyone should go out and read that book.
It's a great one. It's called Influence, right? Influence, yep.
Yeah. Got to go and read that book because this is the way to communicate. And when I teach AI to new people, I'm teaching them communication skills.
It's understanding what you're trying to communicate. It's communicating it clearly. It's making sure the other person understood.
And so I love having his set of... persuasion skills applied to communicating with AI. I just genuinely love it.
So this is something that I hadn't seen until I was researching this. I just went down a rabbit hole, Adam, and this is one of the reasons why I was late to meet and record with you today. I was like, who is this guy?
What is he doing? It turns out, apparently, there's a Wharton study that's co -written by Dan Shapiro and Robert Cialdini that tested these skills, these principles on GPT -40 mini and across 28... thousand conversations persuasion was more persuasion more than doubled compliance it's just like anyway i love that he did this let's move on to the next one then i'm gonna go to my like like agents .md file after this and like add a layer of all these like tildini principles on top what what a good idea All right, number five this week is trading agents.
It's got 118 ,000 stars. This is where agents don't just debate but help you make decisions for trading, but it is all simulated trading. Adam, I've seen a lot of people say, I gave my AI agent $100, $1 ,000, $100 ,000, and here's how it performed when it traded, and they analyze it.
I always thought those people really gave it money. If you look at the details of it, if you really watch the video, you'll see at some point they say, it's all fake money just to see how it works. It's all simulated.
This is the kind of tool that they're using. People are just circling around to see if they could figure out how they can get agents to trade for them, and this is incredibly popular for playing with it without losing any money and understanding the philosophy. I think it's really fun to play with.
I'd be a little scared of playing in this with real money, simply because if this is what you can do with open source and as an individual, what do you think the actual trading desks are doing? How much more advanced? And that's your competition.
That's who you're trading against. That said, I've played with these types of tools before, Andrew, and there is no worse feeling than making fake money in one of them. You put $1 ,000 in, three days later, your fake balance is up at $3 ,000.
You're like, No, I could have had that for real. So I don't know that I'd go out and recommend everybody go start trading with AI, but it does suck to make money doing it, not for real.
Next, number six for the week is Andre Karpathy Skills. It is actually created by Multica, not by Andre Karpathy, the famous AI, what is it, engineer? They basically took his principles.
They turned it into a skill. It's got about a quarter million stars, incredibly popular. I think it's a great idea for Multiket to have created this.
I don't know. If I need this any more than I need any of the other Karpathy skills, what do you think? Yeah, this is kind of interesting.
Actually, the thing that just popped out to me is the last changes all seem to be from six to nine months ago. And yeah, six months ago is the last commit. And so honestly, this is actually a good call out.
These might be great skills and they might have been great six months ago. The problem is these skills really have to be designed for the level of models that we're talking to today. And all of us are talking to...
you know, the newest Opus or Sol. And the skills just won't work as well if they're not updated for that. So even if you want to go and download these, you should go in and say, well, what are the current conventions?
What are the models currently need? Because the input that they used to need isn't the same as what they need today. And I'd want to make sure to update them.
But that said, I'm kind of interested also in going back and looking at Multica. I played with it a lot a couple months ago when we covered it. And I liked the idea, but it didn't quite deliver.
But with the pace of AI development, like if he's still out there building on it, I bet it is. significantly better now. I'd love to go, yeah, active, very active development six hours ago.
So I'd be curious to go and try Multica again, because I liked the idea of kind of co -working with agents through a Kanban board. I've come to the conclusion that none of these work well enough because the agents don't stay disciplined about using it. So unless the agents are built in and one with it, they stray, don't you think?
Would you believe me if I said I have a tool? that can persuade your AI to follow your directives. I believe it said from 33 % to 77%.
It's not 100%. When it comes to task management, you need it. I'll tell you one thing that Claude's Duet is doing well.
Claude came out with Duet. It is, I think, the worst of all the AI agents, worse than DOTS from OpenAI, worse than... by far, than Grokbot from Cursor.
But one thing that they did was they have a to -do list for your agent, and it's built in. And if you say do something, it will put it on the to -do list. I believe it even does it proactively.
And so then there's a discipline. Here, for some reason, they inevitably stray. And I don't know why they can't stay with this, but it's a real problem.
And yes, you can find ways to arm wrestle them. But if you're arm wrestling them into doing a key part of the job, inevitably they're going to get out of the arm wrestle. so they're going to get out of the move.
Okay, that's my take on all this. By the way, Zapier is a sponsor. One thing that I talk about all the time with Zapier is that my GrokBot usage keeps going up, and every Saturday it says you're at like 98 % of usage, and I don't know how to pay them for more.
I've literally sent messages. to Scott at Cursor. I want to pay you more.
Can't do it. It doesn't matter. I don't want to burn through it.
I created a whole agent called Zapier. What it does is it goes into Zapier. It goes into all my routines and says, how can a Zap make it cheaper?
And if I like it, I say, go ahead and do it. And if I don't like it, I let it go. Let me show you also that Zapier not only does that really well, which is turn these products.
Turn these products, these workflows, into code that is cheaper than your agent. They also have a free forever plan, which many people have been asking me about. You can absolutely be on their free plan forever and never pay them.
But I'd love it if you used my URL and told them that you found them for me. Next, taste skills. A skill that gives coding agents an art director.
This has been really popular. I don't know why this week. We're going to see yet another one in the top 10 list.
Do you use this? I'm religious about these. It's so easy to just take the URL from GitHub and say, how would you make this better using this skill?
I love this. Do you use it? I think it's cool.
I don't really use any of these design -oriented skills, though. I much prefer to do my own definition of taste or go and find screenshots from the web and just communicate that. to my agent instead of using the skill.
Now, that's usually because when I use one of these skills, they're too formulaic. And so I would be curious if, you know, if this one or at some point they're not going to be formulaic and it might be now. But also I see a bunch of kind of names here that are listing.
Here's a brutalist design. Here's a minimalist design. It's like, OK, so I kind of get what it's going to produce, right?
Whenever they say ban M -dashes, I think this is like an old thing. We can now allow M -dashes in. I feel like they're missing it.
I still love these things because I need new ideas. It doesn't burn any of my tokens to give them three of these. I would like before publishing a site, I'm on my phone.
I go, here are the top three from last week's GitHub show. Just create a version from each one. And sometimes I go, I really like taste skill.
Have it work with me to create a better version and ask me a bunch of questions. And then I end up with a better design than I would have otherwise. And sometimes it creates something in me.
I can see that. Sometimes I trash it. It seems like a fun variant of the grill me skill from Matt, right?
I hadn't thought that you were going to say that. But with the design flavor, maybe. Right.
I need inspiration. I need somebody to come and give me even a bad idea for me to know what a good idea is. Okay, next.
Oh, by the way, I'm not going to say any more. I'd love to. to 100 ,000 by the end of the year.
If you want to subscribe, go ahead and subscribe. Let's move on to the next. Okay, skills.
14 skills that teach your agents interface design. I'm going to show you what I like about this one. Where's that tweet?
There. This is the kind of thing that I think it does especially well. Let's you go into a device and show it in a way that's easy for people to understand.
I also really like that it has a focus on accessibility. I think that accessibility might have been a pain in the butt before, but now it's becoming easy if you have a collaborator like this tool. You can make it accessible for more people.
Yeah, I think that's pretty good. I also think that these AI tools like Codex and Cloud Code, they're pretty bad at interface design still. We all experience when we're building, there's too many buttons.
There's too much text around the buttons. It's not intuitive for a human to use. But it checked all the boxes of I built all this stuff.
So layering on like this. At some point, agents are going to have this knowledge and understanding. But we are not there yet.
I don't expect it by the end of the year. So having something like this that can. Not just think about the technical accessibility.
Every image should have alt text behind it. You should make sure there's page hierarchy. But even just thinking of how would someone that's reading this page that has full sight, they can hear everything, how would they use it?
What would a human actually need? And again, the agents are still famously bad at this. So I appreciate it.
And I appreciate that it's going to help you bake in the full accessibility kit too so that screen readers and everything can use the sites you make. Of course, if you are seeing me click through all this too fast, there's a link below where you can get it. Yes, I do ask for an email address, but you can put a fake one in.
I'm just trying to connect with you, but I'm not trying to force you. So put a fake one in, and you'll immediately get this anyway. Okay, T3 code.
I was going to start telling you about it. I'm using this through my agent, my grokbot, which I thought was weird, but apparently there are people who are doing even weirder. He did this video.
This is Theo. This is what, a coding harness that allows you to connect Claude, Cursor, Codex, all these models into one, right? Yeah, there's a lot of terminology here that's all changing.
This would be called a meta harness. So the harness is Codex or is Claude Code or Cursor, any of those others. This is a meta harness.
It's not replacing those other ones. When you're in T3 code. You are telling it on this project, I want to create a new chat using this model and this other harness.
And you can select and connect all of the existing harnesses you're already paying for, connect them in here. And then T3 brings all of that into one really nice like GUI interface. And what I love about it is for those of us that are running on multiple machines, you can connect multiple computers to get it together.
You can connect your phone. There's a web version too. from anywhere I can run AI on any of my machines.
And every single one of those computers has the same list of projects, the same list of chats, because they can all see each other and talk to each other. So this is, I mean, it is like, if you're frustrated with... multiple tabs if you're frustrated with trying to use the terminal to manage a bunch of projects and different agents all at the same time like this is clearly the meta harness for you it's the best of the best right now i use it every day i've got three tasks running right now behind this behind behind our video I just wouldn't have enough good things to say about it.
I'm clearly not using it to your level. I have a few things here that stood out for me about his video that I will, of course, link to. He's showing all the different models that now you can work with.
I like how he shows how you can have multiple threads here. No, this is where he's showing. Anyway, you can have multiple threads on the left, but also that he goes, listen, I'm not very opinionated on this.
I want you to be able to do whatever you want and customize this to any way that you like. But one thing that I do like is threads. to be more like tasks that you can settle them.
So you can ask for something. And then when it's there, you're nodding. You've seen this.
You can settle it and go. I like how you can remote control on other computers. I like the example that he gave of a person who never uses it directly, but uses instinct, the AI agent that sits in.
iMessage, and you can only access on iMessage, and is using them to have, I forget how many projects and what he's working on, but the fact that he is exclusively this person using instinct from iMessage to code is an interesting model. The fact that you can get a thread ID and then refer another thread back to that thread ID, that solves a problem that I didn't even know I had that others don't even think about.
That there's a mobile phone simulator that I don't fully understand, but that you can simulate an iPhone in this. Okay, that you can send a prompt now. and say when my limits come back at me, you know, like my limits are going to reset tomorrow, then run that problem.
Like all these things are super well thought out. The part that I don't understand is how the hell does this guy – I guess I say heck because I want as many views as possible from YouTube. How does he only have only 32 ,000 stars on this?
I bet you his skill set is going to have more stars in this. I don't know why. Yeah, that is weird.
Right? Only 30 – I mean there's got to be way more than 32 ,000 of us. I know a bunch of people that are using this.
I do too. But honestly, I probably didn't start it, and I use the thing every day. I'm busy.
I got work to do. I'm running my agents. We're trying to sling businesses, make money.
Did I start it? Maybe I didn't start it too. Yeah, look, I didn't even start it either.
You're part of the problem, Andrew. Start being part of the solution. I think it just goes to show that sometimes you really need what I did at the beginning of this, which is find what people are talking about and then put that out there.
Sometimes it's not just the stars. It's the stuff that's buzzing, that people are excited about. On T3, my friends have been absolutely loving.
And for some reason, people go into it going, I didn't think I would like it. Maybe because it comes from a YouTuber. I thought it was just blah, blah, blah.
And then I love it. OK. Or, Andrew, people tried it six months ago.
And it was terrible six months ago. Was it? I tried it too.
Oh, yeah. And I wanted to love it. And then I moved back off of it and switched to some other tools.
And three or four months ago, it's like, I'll go try again. I tried it again. And boom, immediately, everything switched over.
So if you tried it back in May and you didn't like it, try it again now. Final one with 33 .9 thousand stars. Knowledge Work plugins from Anthropic.
This one we'd seen before, but people, this is accessible and it's by Anthropic, so they know it. It basically is a set of skills to help you with productivity, help you with sales. It's useful.
It's interesting. I don't know what else we could even add to it, except it's here. And if you haven't tried it, go look at it.
What do you think? Yeah, I think you nailed it.
Oh, man. I don't think I've told you about this one yet. We built an app for ourselves that vacuums up all the data about our business.
It's a context layer app. So what it does is it looks and it pulls in every calendar event, every email, every Slack message, every Notion document, every Google document, every call transfer, everything. It pulls it all in and it mines it for mentions of people or companies because we invest in companies.
We like to find entrepreneurs. And so now we have this one context layer that's saying, well, If I mention the name of an entrepreneur on a call with you, that's now going to be linked to that person's profile.
And I can generate like a full dossier of like, what's our opinion of this person? What did Adam mention in Slack? Whatever.
It's all in this one spot. And my God, Andrew, it took us five, maybe 10 hours to build this thing. And it totally works.
And it's running. And it's only going to cost about $100 a month in compute. And I imagine that's going to go down as these models keep getting better and faster.
So I am blown away by this thing that we made. Before I say mine, will you freaking people stop asking what have you built? Like how many people in the comments can say what have you built?
Adam's building stuff all the time. And meanwhile, his job is not to build for you. His job is to build, to invest in entrepreneurs and to help those people out.
These are some of the things he's built. You can say it about me, Andrew. What have you built?
I've shown you how I've used a lot of these tools in the past, and I see the stats. People say they want to see me use it. They get bored to tears, and they clearly will fast forward through the moment when Andrew starts to show a thing.
So I stopped showing you what we do. All right, let me tell you one thing that I have done with AI that I freaking love. This is way simpler than yours.
I now, we're using Riverside to record. I have taken just the Riverside link that I get at the end of one of these meetings, and I give it to ChatGPT, and then I give it links to my other YouTube shows that are similar, and I go in that style, edit this freaking video for me, and it does it. It does it within, I would say, sometimes even, I can't predict why it sometimes takes longer than others.
30 minutes sometimes to do it for the first pass. And then I ask for another pass. And the little skill that I've created for myself has had to teach the agent to be prepared for Andrew's feedback so it doesn't have to rerun the whole thing.
It all happens on my computer. It hardly uses any tokens at all. I might even go back from the $100 a month plan on OpenAI to $20 because I hardly use it even though I do these types of videos all the time.
And I don't tell you people which ones I do because I don't think about it. But now when I have like a 4 p .m. or 5 p .m.
video that I record, I give it to Chachi. and it edits it, and it is up on the site. It is a beautiful thing to see.
And now Riverside is actually embedding that power into their software directly. Okay, that's what I've done. If you like these top 10 repos for this week, there's a link to another set that I think you're going to love, and we'll see you in that one.
The Hook
The bait, then the rug-pull.
The title promises three things at once: a way to copy any app, Claude doing your SEO, and a ranked list. The video delivers in that order, opening on a bonus pick that hosts open-source clones of paid apps for about five dollars a month before counting down the week's ten most-starred repos for agent builders.
Frameworks
Named ideas worth stealing.
03:03concept
Skill realism
Install fewer skills, read every one you install, and customize them to your workflow instead of running defaults.
Steal fora team's rules for adopting agent skills
08:31model
Free data first, paid data for volume
Crawl your own site and use free Google APIs to fix what you have
Layer paid keyword data to learn what people actually search for
Use competitor ranking data to find where volume exists
Free crawl data covers technical and on-page SEO; demand research needs paid search data.
Steal fordeciding when an SEO tool subscription is worth it
14:35list
Cialdini principles applied to agents
Authority
Commitment
Scarcity
Social proof
Unity
Pressure tests
Instruction files written with persuasion principles, such as getting the agent to commit before acting, raise compliance.
Steal foran AGENTS.md or CLAUDE.md rules file
18:59list
Karpathy's four coding principles
Think before coding
Simplicity first
Surgical changes
Goal-driven execution
A rules file that targets wrong assumptions, overcomplication, unrelated edits and unverifiable work.
Steal fora baseline coding-agent instruction file
CTA Breakdown
How they asked for the click.
VERBAL ASK
34:18next-video
“If you like these top 10 repos for this week, there's a link to another set that I think you're going to love, and we'll see you in that one.”
Soft end-screen handoff; the stronger asks came mid-video as two sponsor reads and a subscribe nudge toward 100,000 subscribers.
Add Modern Creator as a preferred source and Google shows you more of our breakdowns in Search, Top Stories, and AI Overviews. It only changes what you see, and you can undo it in your Google settings anytime.
Add to Preferred SourcesOpens your Google source preferences with us pre-loaded. Tick the box and you're done.
Two hosts count down the week's ten hottest GitHub repos, from a security-scanned Claude plugin index to a self-hosted finance tracker wired to 2,500 banks.
A weekly roundup episode where the host and a Zapier guest run through ten open-source repos built to pair with a coding agent, from screenshot-to-code to a proxy that cuts terminal noise by up to 90 percent.
Ben Senescu built OpenSEO to kill his own Semrush bill, wired it straight into Claude, and now watches open-source contributors fork it into CRMs and content tools he never planned to build.
Andrew Warner and Adam run down this week's most-starred GitHub repos — three AI design tools, a parallel-agent "tokenmaxxing" rig, a video-comprehension skill — and pull the geophysicist who built the #1 trending AI job-search framework on camera to explain how it got him hired.
00:11
00:22
00:37
00:52
01:10
01:36
01:51
02:28
02:37
02:52
03:12
03:32
04:06
04:21
04:41
05:06
05:30
06:21
06:47
08:15
08:51
10:03
10:24
10:35
11:37
12:19
12:59
13:38
14:41
14:52
15:29
16:10
16:50
17:56
18:23
19:36
20:12
20:43
21:07
21:31
21:52
22:10
22:57
24:16
25:28
26:02
26:50
27:10
27:54
29:03
30:11
31:00
31:09
31:19
31:34
31:46
32:29
33:23
33:45
34:16
34:19
34:21
34:23
18:45
21:02
13:57
19:02
37:04
39:53