33:27The Next New Thing · Interview
8 Playbooks You Can Try
A host and a product-marketing veteran watch founder clips and debate which of eight AI-era business models actually work — and why.
June 17thThe argument in one line.
The most useful GitHub repos right now are not code libraries — they are packaged skill sets that let a solo builder do the work of a marketer, product manager, and senior engineer without hiring any of them.
Read if. Skip if.
READ IF YOU ARE…
- You are vibe-coding or building solo and want AI to fill in the professional gaps you don't have — marketing, PM rigor, engineering discipline.
- You use AI agents and want to audit the skills you've installed for hidden risks before they touch your data.
- You're paying for CapCut or a streaming subscription and want free, self-hostable alternatives.
- You build agents for others and need real system prompt examples to study structure, tone, and formatting.
- You want a single MCP endpoint that gives any agent access to 8,000+ tools without writing glue code.
SKIP IF…
- You only work in Python data science or backend systems — most of the agent-skill repos target Claude Code / coding agents.
- You are already deeply embedded in Docker on non-Mac infrastructure — the Apple container repo is Mac-only.
The full version, fast.
Every repo here solves a 'human gap' for solo builders: the marketing you don't do, the PM rigor you skip, the engineering process your agent ignores, the costs you don't track. The three hidden gems are a marketing-skills pack that turns an agent into an SEO and copywriting team, a personal-skills repo with a stateful /teach command that learns alongside you, and OpenCut — an open-source CapCut alternative with 57K stars. The top 10 add Apple's native Mac container tool, NVIDIA's skill security scanner, Zapier MCP as a single-URL toolbox for 8,000+ services, a token-compression layer that cuts costs 60-95%, and a leaked collection of real system prompts from every major AI product that builders should study like textbooks.
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.
Create a free account →Where the time goes.
00:00 – 00:36
01 · Intro / teaser
Hook montage: free CapCut, antivirus for AI skills, free global TV, Fable 5 leak. Sponsor mention (Zapier).
00:36 – 00:45
02 · Hidden Gems intro
Andrew frames three under-the-radar repos worth knowing before the top-10 countdown.
00:45 – 02:15
03 · coreyhaines31/marketingskills
Marketing skills repo by Corey Haynes — SEO, cold email, social, onboarding, ad copy. Tested with the author. Turns marketing into an engineering exercise.
02:15 – 05:06
04 · mattpocock/skills — /teach command
Matt Pocock's personal skills repo highlighted for its stateful /teach command: creates HTML lesson modules, gives quizzes, updates as learner progresses. Shown teaching Rubik's cube notation.
05:06 – 06:54
05 · OpenCut — free CapCut alternative
OpenCut: open-source browser-based video editor, 57K stars, 6,200 forks. Runs in-browser, MCP server for AI agents. Not quite top-10 tier yet but a real CapCut replacement.
06:54 – 09:09
06 · #1 — apple/container
Apple's native Mac container tool announced at WWDC26. Runs Linux containers and lightweight VMs on Apple Silicon without Docker Desktop. Written in Swift.
09:09 – 12:27
07 · #2 — NVIDIA/SkillSpector
Security scanner for AI agent skills. Finds prompt injection, credential theft, supply-chain risks. Rated line-by-line. Recommended especially for skill marketplace operators.
12:27 – 14:51
08 · #3 — addyosmani/agent-skills
Addy Osmani's production-grade engineering skills for AI coding agents. 7-phase lifecycle: define, plan, build, test, review, verify, ship. Includes 'extract' adversarial reviewer.
14:51 – 15:45
09 · Zapier MCP (sponsor segment)
One MCP URL exposes 8,000+ tools — Gmail, Calendar, Asana, Notion. No code, no glue. Sponsor but genuinely used by host.
15:45 – 18:00
10 · #4 — phuryn/pm-skills
PM skills that add rigor to vague AI output — benefit-based headlines, clearer requirements. Tested live: rewrote the show's own website headlines from '5/10' to '10/10' with one prompt.
18:00 – 19:39
11 · #5 — iptv-org/iptv
Community-maintained M3U playlist of 125K+ free public IPTV channels worldwide. Paste into VLC. No subscription, no account.
19:39 – 21:00
12 · #6 — Panniantong/Agent-Reach
Gives AI agents the ability to read sites that block scraping — Twitter, YouTube, paywalled content. Plug directly into any agent.
21:00 – 24:27
13 · #7 — chopratejas/headroom
Token compression layer for tool outputs and logs. Cuts 60-95% of tokens before they reach the LLM. Can increase call count when over-filtering. Built by a Netflix engineer, went viral after Fable launch.
24:27 – 27:00
14 · #8 — mvanhorn/last30days-skill
Claude Code skill by Matt Van Horn. Searches GitHub, YouTube, Hacker News, X, Reddit, Polymarket for what's happening right now on any topic. Returns only current results, not 2-month-old takes.
27:00 – 28:57
15 · #9 — music-assistant/server
Home Assistant companion that bridges smart-home audio to Spotify, Tidal, YouTube Music, local files, and more. Solves the one gap that keeps smart-home users tethered to Alexa.
28:57 – 33:36
16 · #10 — asgeirtj/system_prompts_leaks
Aggregated system prompts from Claude, ChatGPT, Cursor, Gemini, Grok, and others. Anthropic releases theirs publicly. The value: studying the formatting, bracketing, and role instructions used by the best AI teams.
33:36 – 33:53
17 · Watch next
Teaser for Grok video creation deep-dive and how to prompt it more effectively.
Lines worth screenshotting.
- Most solopreneurs are product-growth centered and completely miss marketing — a skills repo can do an 8 out of 10 job on SEO, cold email, and ads with no money and minimal effort.
- Every AI skill you install is code that can download files, run processes, and read your emails — treat them as untrusted code, not safe plugins.
- OpenCut has 57K stars and 6,200 forks — fork count is the real signal of utility, not just stars.
- Apple's native container tool at WWDC26 may finally make Docker removable from Mac development machines.
- The /teach skill is stateful — it remembers what you've learned, updates modules as you go, and tailors lessons to what you already know rather than serving a fixed curriculum.
- Agent Reach exists because platforms block bots — and sometimes you want a bot scraping your own site to get your own data out.
- Headroom cuts token costs 60-95% on tool outputs, but can increase total calls when agents need to re-request filtered data.
- The last30days-skill is a better answer to 'how do I stay current on AI' than any newsletter — it pulls only what's being discussed right now on Reddit, Hacker News, X, and GitHub.
- Anthropic publicly releases Claude's system prompts — there is no 'leak.' What this repo does is aggregate prompts from every major AI product so builders can study the craft.
- PM skills and marketing skills are mirrors — engineers need the marketing one; product managers need the engineering one. They fill the gap you naturally have.
- Music Assistant solves the one reason Home Assistant users still use Alexa — it bridges open smart-home platforms to closed music services like Spotify, Tidal, and YouTube Music.
- System prompts from Cursor, Claude, and ChatGPT are the best curriculum for writing your own — study their formatting, bracketing, and role-framing conventions before writing your first custom prompt.
- The Zapier MCP endpoint gives any agent one URL and immediate access to Gmail, Calendar, Notion, Asana — no per-integration setup, no code.
Thirteen repos that close the gaps solo builders ignore.
WHAT TO LEARN
The most useful GitHub repos right now are not libraries to import — they are packaged professional workflows that let a solo builder operate like a team.
03coreyhaines31/marketingskills
- A marketing skills repo can do an 8 out of 10 job on SEO, cold email, social, and onboarding with no money and minimal prompting — the gap isn't capability, it's knowing which skills to install.
- It turns marketing into an engineering exercise: give it a goal (build me onboarding flows, build me SEO sites) and it returns structured output, not vague advice.
04mattpocock/skills — /teach
- The /teach command is stateful: it remembers your progress, updates modules as you learn, and adapts to what you already know — the opposite of a fixed online course.
- Use it on your own repositories to understand why your agent made the architectural decisions it did — so you make better decisions the next time.
05OpenCut
- OpenCut has 57K stars and 6,200 forks; fork count is the real indicator of utility because each fork represents someone who cared enough to modify it, not just bookmark it.
- Open-source video editors are customizable with agents — fork it, drop the repo in Claude, and ask for the specific behavior CapCut doesn't give you.
06apple/container
- Apple's native Mac container tool runs Linux VMs on Apple Silicon with far less memory overhead and fewer crashes than Docker Desktop.
- Introduced at WWDC26, it extends the containerization framework Apple shipped the previous year — not a concept, an actual tool in the repo right now.
07NVIDIA/SkillSpector
- Every AI skill you install is code that can read files, download things, and access whatever tools you've granted your agent — treat them as untrusted code, not safe plugins.
- If you run a skills marketplace, scan every upload with a tool like SkillSpector before hosting it — you have a responsibility to the people who download from you.
08addyosmani/agent-skills
- Addy Osmani's agent-skills enforces a 7-phase engineering lifecycle so a coding agent doesn't skip steps a senior engineer would treat as mandatory.
- The extract skill spawns an adversarial second agent with clean context to attack the first agent's conclusion — the reviewer never sees the original answer, preventing anchoring bias.
10phuryn/pm-skills
- PM skills rewrite vague technical descriptions into benefit-based headlines — the live demo turned a 5/10 title into a 10/10 title with one short prompt.
- PM skills and marketing skills are mirrors: engineers who vibe-code need the marketing pack; product managers need the engineering one.
11iptv-org/iptv
- The iptv-org/iptv repo is a single M3U playlist of 125K+ free public TV channels worldwide — paste it into VLC and cut a streaming subscription before paying for another one.
12Agent-Reach
- Agent Reach exists because you sometimes want a bot on your own site — to extract your data from a login-gated interface without manual copy-pasting.
13headroom
- Headroom cuts tool output tokens by 60-95% before they reach the LLM, but can increase total token spend if the agent has to re-request filtered data — measure before assuming it saves money.
14last30days-skill
- The last30days-skill solves the AI information overload problem better than any newsletter by pulling only what's being discussed right now on Reddit, Hacker News, X, and GitHub for any query.
- Use it for competitive comparisons where recency matters — which AI agent is winning right now, not which one was winning two months ago.
15music-assistant/server
- Music Assistant bridges Home Assistant's open smart-home platform to Spotify, Tidal, YouTube Music, and local files — solving the one reason smart-home users still need Alexa.
16asgeirtj/system_prompts_leaks
- Anthropic publicly releases Claude's system prompts — the asgeirtj repo aggregates prompts from every major AI product so builders can study real formatting, role framing, and instruction patterns.
- Study the bracketing, named roles, and explicit capability lists the top AI teams use — then build a system prompt for your own agent that looks like theirs, not a blank slate.
Terms worth knowing.
- Agent skill / slash command
- A text file (usually Markdown) that gives an AI coding agent a defined process to follow. Installed like a plugin, it runs as code and can access your file system and internet — which is why scanning them matters.
- MCP (Model Context Protocol)
- A standard protocol for connecting AI agents to external tools and data sources. One MCP endpoint can expose hundreds of services to any compatible agent.
- IPTV / M3U playlist
- Internet Protocol Television delivered as a .m3u file — a plain-text list of stream URLs. Paste it into VLC or any IPTV player and the channels appear automatically.
- Home Assistant
- An open-source smart-home platform that can control devices across every major ecosystem. Runs locally, no cloud required.
- System prompt
- Hidden instructions sent to an AI model before the user's first message. Defines the model's name, role, tone, restrictions, and available tools. Every major AI product has one running invisibly behind the chat interface.
- Token compression / headroom
- A middleware layer that sits between tool outputs and the language model, trimming irrelevant data before it consumes context window tokens. Reduces cost when tool outputs are verbose; can increase call volume when it over-filters.
- Fork count
- On GitHub, the number of times a repo has been copied and modified independently. A high fork count signals that people found the project useful enough to build on, not just bookmark.
Things they pointed at.
Lines you could clip.
23:00
“If you can't write the claim compactly, you have a vibe, not a decision.”
02:00
“With a set of skills like this, you could do probably an eight out of ten job on all of those other things that you need to make the business successful with really not a lot of effort and basically no money.”
11:20
“We see a skill, we install it, we don't know if there's a problem in it. We just go for it and YOLO it.”
Word for word.
Read-along
Don't just watch it. Burn it in.
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.
metaphoranalogy
00:00I've got a free version of CapCut that you can install yourself and have work beautifully. If you're installing skills, you've gotta see the antivirus that NVIDIA created. It will save your computer.
00:09Everyone's trying to figure out how to watch soccer for free? Well, guess what? There are television channels all over the world for free on the Internet.
00:15But how are you gonna find them? Well, I've got a repo that will put them on your computer like that. And is there a leak of Fable five?
00:21Tune in. I've got the top 10 repos of the week and so much more. Chapter links below and a full download of everything I'm showing you in the description.
00:29Presented by Zapier, the AI automation company. Alright. Adam, you're back.
00:34We got a whole bunch of repos this week, but this week I wanna start off with something different. I wanna start with three hidden gems that I think everyone needs to know about, and the first one is this marketing skills repo. I think it's freaking phenomenal.
00:48It is built by Corey Haynes. He is someone who's been a marketer for years who basically took all the things that he does to grow a business, and he put them into this set of skills.
00:58I tested it with him in an interview. They work so simply, and they're so freaking effective, and he keeps improving them. And I really wanna draw attention to them because I think that too many people who are looking at GitHub are also not thinking about marketing, and they're missing this.
01:13And that's why I wanted to highlight it. Man, I I couldn't agree more. Like, a lot of these,
01:17uh, like, solopreneurs, like, they're building awesome products. They're very product growth centered, and, oh, I could change this.
01:23I could build this. And what they're missing often, if if you're one person in a business, is you don't know how to run ads. You don't know how to improve your SEO.
01:32You don't know how to do cold emails to your clients. And the reason that you could be this, like, solopreneur now is with a set of skills like this, you could do probably an eight out of 10 job on all of those other things that you need to make the business successful with really not a lot of effort and basically no money.
01:51And it it turns marketing into an engineering exercise that just works. You can tell it, build me SEO sites, build me onboarding flows, and I tested it with him. It works beautifully.
02:02And here you can see this is just from Thursday. He says, uploaded slash social, so he also has social marketing. Instead of post more, it asks which 10 posts should you comment on today.
02:11You it maintains a list of sources. Really effective. Let's look at the next one.
02:15This is from a past, uh, past person who had his skills repo be one of the top 10. We've talked about him in the past. We blew past one of his skills called, uh, teach.
02:25Let me show you this in his repo. You can see here.
02:30This is a collection of skills that he uses to make himself into a better engineer, and so you see the engineering skills here to make himself into a better person. So he's got his personal ones, and also to increase productivity. And there's one here called teach that he tweeted about that I wanna I wanna just highlight here.
02:46And I have a note for myself, um, to go to three minutes and forty seconds where you can see this is what it looks like. And this is really the core of what makes this such a good learning experience. And here's my first lesson, anatomy notation.
02:59And and it creates HTML docs with each lesson as you just showed. Cross. And this is him using it to learn how to use a Rubik's cube.
03:07Basically teaches you just what you need to know at that moment. It gives you diagrams. It gives you very simple explainers.
03:14It gives you callouts, and it gives you quizzes too. One thing I find really important whenever you're teaching anything is to develop a feedback loop. And quizzes are okay at this.
03:24They're basically good if you can't find any richer feedback loop. Again, we zoom down this, we've got notations, we've got traps.
03:31We've got Okay. And he's just showing how he learned the Rubik's cube. And one of the reasons why he has this here is to show that there's some skills that he has that will remember what you've done stateful versus stateless, and this is an example of that.
03:45Any thoughts on this?
03:47I I love the learning especially because it creates multiple modules there. Mhmm.
03:52So you can actually go through one of the modules, you know, one of those HTML files, And then you could go back to the agent and say, hey. I learned this.
03:59I think this. I have these questions. And the rest of the modules could actually change them as a result.
04:05You can update this stuff as it's going. It's like, oh, I've learned this. Or I I'm this is cool, but I'd like to learn about this use case instead.
04:13And the learning is, like, really dynamic then compared to you know, you could find a course on how to solve a Rubik's cube online, but it's not gonna be tailored to what you already know, and it's not gonna keep up with you as you learn. It's pretty formulaic.
04:28So I I love this sort of thing. And I think you should be using it if you are vibe coding out there. You know, I don't I don't think people put enough focus on, well, why did my agent choose these technologies?
04:38Why did it choose this architecture? What are the options? I would be using this slash teach skill on my own repositories and on my own apps that I'm deploying to say, hey.
04:48Tell me about how these work, how they're created. And, like, wow. When I come back to it to work on it again, like, I'm gonna know exactly how the things work.
04:56I'm gonna make better decisions the next time around because I'm learning from my own successes and failures as as I build out apps. The more I get to know his work, the more I like him, and I like his work, and it's available for everyone there.
05:08Last hidden gem. This is an open, uh, a CapCut alternative.
05:15I like CapCut a lot because it simplifies video editing and it makes it powerful. The problem with CapCut, not a huge issue for me, but it is a pain in the neck.
05:26They charge you for stuff or they try to promote their own things in it. I say not a big issue for me because I live in video. But if you're doing video editing for like family videos, if you're doing video editing yourself in a way that you may not wanna pay CapCut or maybe you just don't want their software on your computer, which actually is a huge issue for me, this is an open source alternative.
05:44It has the features that you need. You can build onto it. You can put it on your device, feel good about it, and it's I'm glad to see it here as one of the hidden gems.
05:53And it's actually it's it's not yet at a level where it could reach a top 10. It's got 57, uh, thousand stars, though, so it is popular.
06:026,200, uh, forks, which you've told me to watch out for to get a sense of real indication and utility.
06:08Yeah. These are cool. I mean, I grew up learning Premiere and After Effects and all those.
06:13And, oh my god. I almost no one is gonna learn those, like, wildly in-depth tools. And you don't need to.
06:20I don't want to. For almost any content that people are creating today, unless it's a feature like film, you really don't need the tools like that.
06:26So I I love this. I love the competition. I love open source because all those forks, Andrew, each one of those are an opportunity for someone to say, oh, I like this tool, but want it a little different.
06:36And we've already talked about this. You can use agents to customize open source tools. So if you're using OpenCut, you're like, oh, I wish it did this instead.
06:44That's great. Fork it. Put that repository in your own computer and ask Claude to make the changes that you want so that it works in the way you like.
06:52That's it's I mean, I love this ecosystem.
06:55Okay. Let's look now at the number one repo of the week.
06:59This is from Apple. Docker desktop has always felt heavy on a Mac. Apple's own tool runs Linux containers and lightweight VMs built for Apple silicon.
07:08This is the GitHub repo. I'll show it in a moment, but first, this is for one of the people at Apple talking about it. One of my personal favorite features announced that WWDC will, I suspect, be a sleeper hit container machines allowing your Mac to run a lightweight persistent Linux environment.
07:22And then someone came in and said, this existed for a while. It's helpful, but it's nowhere near as flexible as something like Docker to which he came back in and replied, container machines are brand new capability this year.
07:39They build on the containers, uh, container support we shipped last year. He goes, wait.
07:43I didn't read your bio. That's cool, though. Yeah.
07:45Developer tools and frameworks at Apple. It's Yeah.
07:49And then he did start to explain it. At WWDC twenty five, we introduced generalized container support with an open and I'd for you to explain what this means.
07:59Uh, with an open source containerization framework and container CLI, container machines introduced at WWDC twenty six extend that enabling a seamless high integrity Linux environment on a Mac.
08:13I'll leave this for people in, uh, the dock, which we'll have a link to in the description. But why are you excited about this? I mean, I have Docker desktop running on my machine.
08:23It crashes a lot. It's slow. It uses a lot of memory.
08:26I would love to get out of it. I count me and the group of people that didn't know about this tool until you introduced it to me here for this show. I was like, oh, man.
08:35Like, I think I'm gonna love this. Because, you know, we use these machines all the time. If you're building apps, each app probably has its own little container where it gets to run-in its environment.
08:46It has its own database, and that's the thing that you deploy out to the cloud when you wanna share it with with actual customers.
08:53And, man, Docker is kinda heavy for that. You're running a bunch of copies of Linux. The tool itself is kinda buggy.
09:00Uh, so I'm very excited to, uh, to try this out and see if, man, if this can't just take Docker off my machine.
09:07Alright. I love when you get excited about discovering a new thing. Alright.
09:11NVIDIA Skills Spectre. I've heard this called basically antivirus for skills.
09:16We see a skill. We install it. We don't know if there's a problem on it.
09:20We just go for it and YOLO it. That's the version of, uh, that's the vision behind this. And you can see here I I can even show how it gets installed on a computer.
09:31You can see here a Nvidia Skills Spectre assessment on Angular where it found a problem that's critical.
09:40This is the report it got. And what do you think of this report here?
09:45First of all, I love it. I I think everybody should be a little bit more skeptical of the skills that they're installing because they don't feel like code, but they're code.
09:56They can download things on your computer. They can run code. They could copy all your emails if they wanted to, if that's what you've given your agent.
10:03And so you should be a little careful when you're downloading all these skills. And you're not gonna read them all.
10:09I'm not gonna read them all. Having something like this installed where you can run it against all of the skills you've installed on your computer.
10:15Like, honestly, I think everyone should pause the video for a moment, download this, run it, install it, see what it what it calls out as issues. Now I think the complaint that's coming up here is that the, you know, this repository maintainer is saying, hey.
10:30It's calling some things out that that maybe aren't real. But there are places to be skeptical. You know?
10:36If it's I almost guarantee you're gonna find something on your computer that this tool says, hey. You should look at this, and at the at least look at it.
10:45Let's take a look quickly at how it works here. So I'm using SkillSpectre.
10:49Inside the GitHub repo, there's this test folder. And inside that, they've got some dangerous skills you can actually run it on to confirm the tool works. So we ran it on these skills.
10:59And with every one of them, it tells you not to install. The higher the score, the more dangerous the skill. And with each test, it doesn't just give you a number.
11:06It shows you the exact line number, the exact location, and the
11:10Okay. Now this is something that you pointed out to me before. You said, well, it's not enough information.
11:16The line number is kind of a pain. I wish that they would say what's there, and then the remediation is not helpful because it's it's cut off.
11:24Yeah. I don't understand that. That does seem to be how it how it works.
11:27My guess here, Andrew, is that there is an interface that you could run this with as well. Or clearly, if you had this output and you fed it back into your agent, the agent could build you a beautiful HTML report of exactly what's wrong. The agent has enough information here to to run with it.
11:44There is a class of person that I think needs to install this now, and there's a lot of people out there building these marketplaces for skills. I'm gonna build a marketplace for marketing skills or sales skills or engineering skills or whatever.
11:57And all those people, if you're hosting a marketplace, please install something like this when a user uploads a skill to share with others. Pass it through a system like this that's gonna rank it.
12:09And if it ranks it as potentially dangerous, you should flag that for review before you encourage other people to download it. I know this is a pretty popular trend now.
12:18Uh, please, like, take a little bit of responsibility and and, you know, scan these be before you host them. I wonder if I should do this even with our team. You know?
12:26If we're starting to share skills, we should be a little bit more careful about what people are sharing.
12:31Okay. Next, agent skills. Most AI coding output is sloppy.
12:36What this is gonna do is it's gonna turn your AI output or at least your AI coder from just an intern into a senior disciplined engineer.
12:49I I don't know. You've taken kind of a skeptical view of these, but also talked about the utility of them. What do you think of agent skills?
12:56Yeah. You know, kinda like the marketing one we looked at a little bit ago, I think this is a great option to fill in a gap. If you're sitting here and you're vibe coding and you're trying to sell this, maybe you're even making some money now Mhmm.
13:09But you don't have an engineering background, going to a set of skills like this and almost just using it blindly and trusting the process, he's got a very specific timeline that he says you move through. You go through define and plan and build and verify and review and ship.
13:25And you know what? If you don't have the background, you might not think to move through all those steps. So just installing a skill pack like this and trusting the process is gonna make you a better vibe coder, a better builder.
13:36And then you know what, Andrew? How about you couple this with that teach skill that you looked at before? And you're like, oh, why do I do this?
13:43And what should I learn from this? Because a regular engineer mostly is isn't gonna need skills like this. They're just gonna do these things by default.
13:49Well, you can learn how to do that too. Go and use the teach skill.
13:53Alright. I call this a report, but I actually built full blown websites, um, to to help guide this.
13:59Extract. Here. Before any nontrivial It's gonna look at one of those.
14:03The biggest skill in the repo. Before any nontrivial decision stands, your agent spawns a second agent with a clean context to attack it. Claim, extract, doubt, reconcile, stop.
14:18Three rounds maximum. The clever part, the reviewer never sees the original conclusion because handing it your answer biases it toward agreement. And my favorite line in the file, if you can't write the claim compactly, .
14:34But does any of this fix why agents actually fail? Because they mostly fail from wrong context, not low intelligence.
14:42That's what the context engineering skill targets.
14:45Okay. And so it's a collection of skills. This video explains some of them.
14:49I'll go on to the next. Oh, the next one is Zapier. One of things that I love about creating things for myself is getting to try lots of different tools.
14:57My headache is constantly that I have to bring all my connections from one tool to the next. And from literally, lately, it's been from one computer to the next. The thing I love about, uh, Zapier MCP is I connect everything into this one place at Zapier.
15:12I give it whatever control I want over all the tools that I have, and then I have one connection, one toolbox essentially that I give Hermes agent, that I give Claude code on this computer, that I give what do I have on that computer? Whatever it is, uh, I get to give it to it, and it just works beautifully.
15:29And it's from Zapier, a company you know, you can trust, and that, um, can become kind of the middleman between your most valuable data. Think about all the things we're giving AI access to between your most valuable data and the tools that you're experimenting with. Go try it for free right now at zapier..com/mcp.
15:45Next.
15:48Oh, I have something good to show about this one. This is PM skills. Generic AI gives product managers text.
15:54This gives them rigor. Let's take a look. Actually, you know what?
15:58You and I were saying, I couldn't find enough to test this on.
16:02So we tested it on our stuff. But before we do, let's just look at the GitHub repo. Is there anything you wanna say about it before I show what what we did with this?
16:10It's the same category. If you're a project man or a product manager and you need engineering support or marketing support, go use those other skill packs that we looked at. If you're an engineer and you're vibe coding, you need that marketing one, and you need this one.
16:22Like, these are here to fill in the gaps that you have naturally. It's not it's not a failure. You know, we're all only gonna know so many things.
16:30These help fill in the gaps. And, honestly, I think this one specifically is a is a pretty good I I like the results out of this package.
16:38Alright. Here's what we did. We basically one of the things that it does is it helps you think through the way you explain things in a more, like, more benefit based.
16:48So I gave it this whole, uh, website that I created to guide our conversation today, I said, how would you change the headlines? What would you do?
16:56And it said, well, look. You've got this Apple container. Before, you explained it as a swift tool that runs Linux containers as VMs on a Mac OS, or this is the way you might have.
17:06After, I would I would say using this skill, say it sorry. Say Docker desktop, but free, native, and built by Apple. So for the NVIDIA Andrew, that's so much better.
17:17Yes. And it's like, we went from a five out of 10 title to a 10 out of 10 title, and we didn't have to learn anything about,
17:24like, how do we do the the marketing of this? How do we do the explanations of this? Like, the skill did a 100% of that.
17:30I mean, I watched you prompt this. It was like one short prompt, and this is what came up with. I'm actually going to give my skill that creates these websites for our weekly talks.
17:39I'm gonna give this to it. This site I don't think this is the way that I exactly explained it. A security scanner for AI agent skills.
17:45That's the NVIDIA Skill Spectre. But I do like how it said it. Scan any agent skill for hidden malware before you install it.
17:51Alright. So all of it really helpful. It does give us a taste of how this works.
17:55Let's move on to the next one. Oh, this one's been so hot because of, uh, because of soccer being so big right now.
18:04We know that there are ways for us to watch television online somewhere, if not in our city, not in our state, not in our country, somewhere around the world.
18:13But how do you even find it? What this does is it it it's basically a collection of all those little channels bundled up in a way that you can give your video player so that it just freaking works in the video player.
18:26And so I just really like it. I like it today. I like it forever.
18:30I I think it's really a beautiful repo. What do you think?
18:34I I love any of these collections of open source things that just, like, make life easier Mhmm. That are finding something free, finding some, like, gem on the Internet.
18:43Now because it's not a full package, it's like you gotta install, you know, one of these IPTV players or VLC or whatever.
18:52You know, they've got a whole list of different things you can play from here. I would if if you want it just to be easy, just download VLC. Right.
18:59There's one for Windows. There's one for Mac. Download VLC, and it's very easy then to just paste in the the the playlist that they have here because it looks pretty overwhelming.
19:10But I'm I'm telling you, this is easy to use, and, uh, it's
19:14not overly technical. It's so great. I had no idea this existed until I saw it just really take off and trend this week.
19:21Um, and they're saying, look. Before you pay for another streaming subscription, here's a trick.
19:26Just go and get it. Put it in VLC. I've got a video here of how to install it and how to use it, but I don't think it makes sense for us to show it now.
19:32I'll have it on the site if anyone wants it. Next, agent reach.
19:38Uh, this is a problem that you and I come across a lot. You ask an agent to read a tweet or a YouTube video, and that's where it goes blind. I know there are other situations, but that's especially the one that hits me every single week.
19:49And so what agent reach does is it gets past their blocks. It makes it easier for you to access it and and use those sites. And whether you're using a personal use or you're building something, this has been really interesting and helpful.
20:01What do you think about about this one?
20:04I mean, yeah, there's this battle on the Internet of bot detection and bot prevention and, hey, sometimes I want a bot. I wanna scrape this site. Even if it's my own site, I wanna be able to scrape it.
20:15Right. I wanna get my data out of a a site where I have to log in to see it. So platforms like this, you know, there's a bunch of different tools you could use.
20:22Some of these tools are meant to control your browser. They're kind of slow, but they give you a lot of control. Some of these tools are meant to be built into scripts and applications.
20:33Some like this are meant to be used by agents. So this one is great because you can just plug it directly into whatever agent you want and immediately give it the tools to start pulling content from protected sites. Yeah.
20:45I love it. I don't I don't have much more to say. It it's kind of everything you need in one package.
20:50I'm glad it exists. I wish it didn't have to. I'd like my agents to just have clear access.
20:55Alright.
20:56This is by a Netflix engineer who decided that he wanted to pay less per token or for token use. And this really took off after Fable came out because Fable was very token hungry and people used it, discovered it. Uh, Fable, of course, was pulled back, but now people still have it.
21:14They still like this. Um, it's still talked about. I'm gonna show you an example, actually.
21:19This is, uh, this is the repo here. I'm gonna show you an example of how one of the cofounders of, uh, Hermes, my favorite agent, tried it.
21:28And the first thing they said was, um, key conclusions. Most of what it would do for Hermes actually is a net greater token cost, but there was one aspect it discovered that would save 60% of tokens used on search file tool, which we will be integrating, and this took off on Twitter, and I'm glad to hear that.
21:47Why why is it that sometimes it's more, sometimes it's less?
21:51Yeah. AI is so frustrating. Right?
21:53It like, sometimes it's backwards. You know, the the tool calls that are made a a tool call sounds weird and technical, but it's really just, oh, if you connected your email, connected to your calendar, your Notion, or whatever, in order for Claude to reach Notion, it uses a tool.
22:09And it says, Notion, give me some data. Well, often, the tools that are built by Notion or Gmail or whoever, they give back too much data.
22:18You know, the agent has requested emails for the last week, but the tool also will return a bunch of metadata or the content of the emails. Maybe you only wanted the subject. So a layer like this in between the two says, oh, I see what that tool was trying to give me from Notion, but I can tell that the agent didn't want all of that information.
22:37So let me cut out the things that are irrelevant so that they don't have to make it all the way to the agent. Well, hold on a minute, though.
22:44What happens if your agent decides, no. I did want that data. Well, now it needs to make another tool call.
22:50So it's not that it's using more tokens per call. It's that if some data that the agent was expecting got filtered before it made it, now it's gonna have to ask again, and it's gonna have to ask more explicitly.
23:02And I bet what they found was a lot of repeated calls looking for Nuance data in different ways. And there's some cost to every single call that goes up if if you, you know, get rid of it.
23:15So I like to give my agent as much context as possible because they're so intelligent that when you give them that context, now I can ask great questions.
23:24I can generate great outputs. But, of course, yeah, man, Fable is expensive. These new models, they're getting more expensive.
23:30They're generating more tokens for the same messages. So if you can cut cost, I I would try it.
23:36I would try and plug something like this, uh, into your into your toolkits.
23:40Uh, but don't, you know, don't forget to check and see if you're actually saving or not. By the way, if you are getting this website after, and there's a link in the description, I have all these links that I used to show in the past, but I found that it slowed down the conversation. But when you're ready to go deep into one of these repos, the links will help you because they're other people's guided experiences and more, uh, nuanced evaluations.
24:02And I spent a lot of time putting that together for you. It's not just AI. It's me, which means you'll sometimes find mistakes in it.
24:07I don't think I have many, but sometimes you do. I do also have this video of, uh, BetterStack doing a test before and after.
24:15I don't think it makes sense to show it here again, not to slow it down, but I think it is interesting to see him specifically try one thing with it and one thing without, one prompt with, one prompt without, and look at the results. And, uh, it saved him tokens. Alright.
24:28Last thirty days skill. This is coming back. This was number one last week.
24:32It is now number eight this week. It is insanely popular on Twitter because Matt Van Horn is so popular on Twitter and because it is so helpful.
24:40Um, what I like about this is that it it understands that when you wanna learn something, you don't want to go and get not always get everything that's on the Internet. You want a few curated sites that you trust, GitHub being one, YouTube being one, Hacker News, uh, x, but only most recent, uh, responses.
25:05And so the example that he gave on one podcast recently was, I wanna send out cold email that gets me responses. What's working today? Go find the people who are showing off on Reddit and give me that.
25:16Another example that I seen him use somewhere was, um, legal questions.
25:22Help me understand what legal questions people have today that are not solved for this area, and he'll get that. Or Open Claw versus Hermes versus Paperclip is a really good one because these just came out. Don't give me what happened even two months ago because two months ago, Open Claw was number one.
25:36Hermes was weird. Now tell me what they are. I think this is great.
25:41You know, Andrew, you know this. I'm I'm the CTO of a venture studio. We own a bunch of companies.
25:47And one of the questions I get the most often from anybody at these companies is how do I stay up to date on AI? Mhmm. And honestly, I don't have a good answer because there's so much content.
26:01You gotta follow this blog and subscribe to this newsletter. And pretty soon, you're just spending all your time reading and watching YouTube, and you're not doing your job. I think this is a pretty I wanna test this out as a way where I can give people a very customized tool and even stream of, hey.
26:18You're interested in marketing tools for SEO and GEO. Well, great. Run this every Monday and and get a custom report that just has what's interesting and useful to you.
26:30Because because until I saw this, I don't I don't feel like I've had a good answer to that. There's not one newsletter I'd recommend. There's not one YouTube channel I'd recommend.
26:39You've gotta be immersed in it, and maybe this is a good way to get immersed.
26:42Will somebody please introduce me to Matt? I want him to come on here. I'd like Matt to come on here with me and Adam.
26:48I'd like him to just do a session just on on the repos that he has. He's been doing incredible work. I don't know why, but I'm glad that he is.
26:55Oh, I also like that he uses Polymarket because that's actually becoming a really insightful source for us. Okay. Alright.
27:01Let's go on to the penultimate one. This is music assistant server. I didn't know what this was.
27:07You and I talked about it. Let's go on GitHub. Tell me what this is, and I'll show the the pages that you that you were referencing.
27:13This is this is a niche cut. There's this platform called Home Assistant
27:19that people use to build their own smart assistants in their home. So you could already get to Alexa or use, you know, the Siri version and control your devices. You could talk to it.
27:30A bunch of people have, uh, have used this open platform called Home Assistant that can connect all smart devices from pretty much every platform. The problem is it's difficult to couple those with tools like Spotify or with Apple Music. Those are very closed ecosystems.
27:46And Home Assistant being an open platform, it's been around a long time. They try and stay pretty neutral on this. And so this Music Assistant I mean, look at how many different sources of music you can couple with this.
27:58I don't know about you, Andrew. Like, I have Alexa at home. And I ask it about the weather, and I ask it to play music.
28:06And Yeah. If I had Home Assistant, I'd struggle to get it to play music. Well, with a thing like this, you can pair all these different sources of music with it.
28:15And even, like, coupling in YouTube Music and saying, oh, I wanna play this song. Which platform do I have rights to listen to this song on? This package kinda takes care of all of that.
28:24So it's a wonderful companion for Home Assistant.
28:27This actually is very compelling because of music. It drives me crazy that my a word I won't say it, so it won't trigger it. My my that's it.
28:34You'll have to bleep me. But it won't play YouTube Music.
28:39And you know what? That's the only thing that's keeping me from using YouTube Music, um, as my main music player because YouTube Music has all the music versus Spotify, which doesn't.
28:50Yep. Okay. Alright.
28:51I got a video here for people who are interested. We're gonna look at the very last one. System prompt leaks.
28:56I'm gonna show you the tweet that is indicative of the kind of thing that people got really psyched about with this. Someone leaked the entire Claude Fable five system prompt, and Anthropic cannot stop it. They cannot stop it, Adam.
29:10A 120,000 characters. What what do we do?
29:13Meanwhile, Adam, you say this is this is not that. What what is this then? It's not it's not some secret that we've uncovered.
29:24First of all, Anthropic releases their system prompts. So we could talk about what a system prompt is, but the but the hype around, oh, no.
29:33They released it. Anthropic releases all these. You can go find it on their website.
29:38Oh, yeah. There you go. You've pulled it up.
29:39Here's the here's the prompt. So I don't know that it's the the concern, but but I am actually curious, Andrew.
29:47I'm not sure that everyone in the audience would know what a system prompt is or why it's so important. Do you know like, how would you define a system prompt?
29:56The only thing so only because I've read this, I I have an understanding. It seems like it's the hidden prompts that go to the model before my prompt.
30:07Like, as I was reading it on on Anthropic site, I think I saw something like oh, here. Child safety. Like, before it says, uh, before it sends anything in, it says Claude can never create romantic or sexual content involving direct minors, etcetera.
30:23So it's like the prompt before my prompt. Am I understanding that right? Yeah.
30:26That's that's exactly right.
30:28And it's actually important to read these. If you're out there, you're vibe coding, you're building agents, you really should go to this site.
30:37Look at the system prompts for the models that you're often using and and learn from them. These are all templates.
30:44When you're building your own custom agent that's talking to GPT five five or whatever, you don't have this system prompt.
30:53But if you go to chatgpt.com or if you go to claud.com, this prompt is running before your very first message.
30:59And it's the opportunity that we have as application developers to inform the model on, here's who I want you to show up as.
31:07Your name is Claude, and you're the Fable five model, and here are the things that you do. Here are the things that you don't do. It's all these hidden instructions.
31:16And if you haven't built a custom set of system instructions for your app, you really need to because the agent is expecting it. The model is expecting this set of instructions in order to tailor all of the output.
31:30So when we build these, we tell the agent about the tools it has. We tell it what role we want it to play. We tell it how we want it to do the work.
31:39And these are literally system prompt examples from the, like, top tier model providers on the planet.
31:48So you should be reading these. You should come back here and run that teach skill against them and say, teach me. How do I write a great system prompt based on this, like, huge repository of system prompts.
32:00I see. So I come to these all the time, Andrew. I'll go to the cursor one, I'll say, how are they doing it?
32:05What tactics are they using? What formatting are they using? Oh, they've got all these, like, weird little, like, brackets and names.
32:12Why are they doing that? These are all things that if you're not incorporating into your system prompts, you should be. Who would you rather listen to about how to write this?
32:20Me or Claude?
32:22This is a great way to go and learn how the Claude team writes their prompts. I see. I thought this I thought you were gonna tell me to understand restrictions behind the model.
32:30Like, here, legal and financial advice. Formatting, the structure It's how are they prompted. Included in here.
32:36What are they not included in here? All of those things, you should be learning. And if, again, if you're building an agent for other people, learn from the best of the best.
32:44And I wanna see a system prompt in your app that looks just like these because these guys know how to do it. I see. Okay.
32:51So it is a collection of how they're telling their agent to talk to me. Like, for example, Claude uses warm tone, treating people with kindness and without making negative assumptions about their judgment or abilities. But it's also here is how we like to prompt.
33:05This is the structure. It's kind of interesting because is is Grock in here?
33:08No. Grok is not under miscellaneous. No.
33:11They're are not. The reason I ask is because people have been showing the videos that they're creating using the new Grok video creation model, and my videos just suck.
33:22And one thing I noticed having watched people's YouTube videos is they are going deeper longer with their prompts than I am. I'm giving it one or two sentences, so I could see that that's one issue, but maybe there's something else to the nuance that's helpful. Alright.
33:35That is the last one. In fact, speaking of that, I've got a list of things that launched in the last week here that are not on GitHub that you wouldn't have paid attention to hidden gems, and one of them is Grock with an understanding of how the video works, and go watch that next. See you there.
The Hook
The bait, then the rug-pull.
Every week, builders are paying for software they could own — and ignoring the GitHub repos that would replace it for free. This week's roundup covers 13 repos across video editing, AI security, token costs, and marketing, with the kind of specificity that makes the list actually useful: what each one does, why the numbers behind it matter, and which gap in your solo-builder stack it fills.
Named ideas worth stealing.
13:20list
7-phase agent engineering lifecycle
- Define
- Plan
- Build
- Test
- Review
- Verify
- Ship
Addy Osmani's agent-skills enforces this sequence on every coding task so the agent doesn't skip steps a senior engineer wouldn't.
Steal forAny vibe-coding workflow — install it and trust the process even before you understand each phase
14:00model
Claim-Extract adversarial review
Before any nontrivial decision, spawn a second agent with clean context to attack the first agent's conclusion. Three rounds max. The reviewer never sees the original answer to avoid anchoring bias.
Steal forHigh-stakes code or architecture decisions where you want a second opinion without bias
How they asked for the click.
VERBAL ASK
14:51link
“Go try it for free right now at zapier.com/mcp”
Host positioned it as personal use case (one MCP URL for all agents across machines), then dropped the URL. Felt authentic because the show genuinely uses it.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
