07:22The Next New Thing · Interview
We ran Fable through 600+ tests. It's amazing
Zapier's Automation Bench ran Claude Fable 5.0 against hundreds of realistic business workflows — here's what the numbers actually mean.
June 9thThe argument in one line.
The most useful open-source AI repos this week all solve one underlying problem: giving your agent cheaper, broader, and more controllable access to the world — whether that is the live web, your local files, or your own compute.
Read if. Skip if.
READ IF YOU ARE…
- You run AI agents and are burning through tokens after adopting Claude Fable — headroom and markitdown directly address this.
- You use NotebookLM and want local model support, multiple speaker voices, or full data privacy.
- You build with Codex and want to connect it to Notion, Asana, Jira, or Gmail without a paid middleware layer.
- You are frustrated that every AI-generated UI looks the same and want a prompt-based design system to fix it.
- You manage markdown notes and want a beautiful, file-system-native, AI-friendly alternative to Obsidian or Notion.
SKIP IF…
- You are not actively building with AI agents — most repos require setup and coding familiarity.
- You need production-ready, maintained SaaS products rather than early-stage GitHub repos.
The full version, fast.
This week's standout GitHub repos attack token bloat and agent blindspots simultaneously. last30days-skill queries X, Reddit, YouTube, HN, and Polymarket in parallel using your existing browser cookies as auth — no API keys required — and returns a ranked brief of what people actually engaged with. headroom compresses JSON, logs, and code by 60-95% before context hits the LLM. open-notebook gives you a self-hosted NotebookLM with local model support, multiple speaker profiles, and no Google dependency. rilable demonstrated that Claude Fable can reproduce a full mobile app from screenshots in two prompts. markitdown, already at 110k stars, converts PDFs and YouTube videos into clean markdown and cuts token usage by up to 70%.
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 · Cold open — three pain points teased
Andrew hooks with markdown overload, slop UI, and blocked agents as the three problems the episode solves.
00:36 – 03:54
02 · last30days-skill — real-time social research agent
#1 repo of the week. Queries X, Reddit, YouTube, HN, Polymarket in parallel using browser cookies; synthesizes a ranked brief of what actually engaged people in the past 30 days.
03:54 – 06:09
03 · headroom — token compression toolkit
Compress tool outputs, JSON, logs, and RAG chunks before they reach the LLM. Claims 60-95% fewer tokens via SmartCrusher, AST-aware compression, and CCR.
06:09 – 09:00
04 · open-notebook — self-hosted NotebookLM
Open-source NotebookLM alternative with local model support, multiple TTS speaker profiles, and full self-hosting. 25k stars.
09:00 – 09:54
05 · Zapier MCP sponsor
One MCP URL giving Claude access to 7,000+ connected apps. Sponsor integration.
09:54 – 10:03
06 · Repos Worth Checking — intro
Transition to the bonus 10 repos segment.
10:03 – 11:15
07 · ian-xiaohei-illustrations — blob character explainer graphics
Codex skill that turns any article into hand-drawn blob-character explainer illustrations using GPT Image 2.
11:15 – 12:54
08 · tolaria — offline markdown knowledge base
File-system-first, Notion-like markdown app that AI coding tools can read/write directly. Open source, offline, private.
12:54 – 15:00
09 · rilable — Claude Fable clones Lovable in 2 prompts
Riley Brown used Claude Fable to clone the Lovable mobile app UI, then had it build Notion inside. Proof of Fable's raw capability.
15:00 – 16:48
10 · pm-skills — 68 PM skills and 42 workflows
Large prompt library for project management: strategy, consulting, interviewing, prioritization. Best used as a steal-and-customize starting point.
16:48 – 19:21
11 · openai/plugins — official Codex plugins
Official plugins for Codex: Figma, Notion, Asana, Gmail, Calendar. Designed for spec-to-implementation workflows.
19:21 – 21:27
12 · Agent-Reach — zero-API-fee social access CLI
CLI giving agents granular access to Twitter, Reddit, YouTube, GitHub, Bilibili, XiaoHongShu via cookies.
21:27 – 23:06
13 · NVIDIA/cosmos — physical AI world models
Open platform of world models for robots, autonomous vehicles, and smart infrastructure.
23:06 – 24:27
14 · GitHub Star Controversies — fake engagement
PhantomStars bot flags fake star campaigns. Several repos covered this week were flagged.
24:27 – 27:18
15 · taste-skill — design taste system for AI-generated UIs
Prompt library steering Claude away from generic AI aesthetics toward brutalist, minimalist, or soft design systems.
27:18 – 29:33
16 · apple/container — macOS Linux VM integration
Apple's official tool for running Linux containers on Apple Silicon. WWDC 2026 announcement; Apple's attempt at a WSL equivalent.
29:33 – 32:21
17 · microsoft/markitdown — any file to clean markdown
110k-star Microsoft tool converting PDFs, Office docs, images, and YouTube videos to markdown. Cuts LLM token usage by 70%, includes MCP server for Claude Desktop.
32:21 – 32:25
18 · Outro — watch this next
Teases next episode on underreported AI releases overshadowed by Fable.
Lines worth screenshotting.
- Browser cookies are already a valid API key — last30days-skill uses yours to bypass X and YouTube paywalls without any developer credentials.
- A 20-page PDF burns 70,000 tokens before you ask a single question; converting it to markdown first drops that by 70%.
- headroom cuts LLM token usage by 60-95% by compressing JSON deduplication, AST-aware code stripping, and BERT-based context pruning before anything hits the model.
- Claude Fable can clone an entire mobile app from screenshots in two prompts — rilable proved this by rebuilding Lovable and then having the clone build Notion.
- open-notebook replaces NotebookLM with local models, multiple speaker profiles, and full self-hosting — the setup cost is higher, but you own the entire pipeline.
- Taste Skill does not give your AI original design taste — it trades one set of cliches for another, but it is a useful starting point for building your own design system.
- GitHub star fraud is detectable: PhantomStars flags repos where suspiciously single-purpose accounts bulk-star on the same day — several of this week's top repos triggered it.
- apple/container is Apple's attempt at a macOS WSL — running Linux VMs with synced file systems on Apple Silicon, announced at WWDC 2026.
- Agent-Reach gives agents granular, zero-API-fee access to Twitter, Reddit, YouTube, GitHub, and several Chinese platforms — more surgical than a full research pipeline.
- markitdown has an MCP server for Claude Desktop that auto-converts every uploaded file to markdown before it hits the model — install once, save tokens forever.
- The best PM prompt libraries are steal-and-customize starting points, not frameworks to adopt wholesale — the skill files matter more than the folder structure.
- open-notebook uses Docker for setup and local Ollama for voices — the quality ceiling is higher than NotebookLM, but the setup floor is higher too.
Ten tools that make your agent smarter for free.
WHAT TO LEARN
The most useful GitHub repos this week all reduce the same tax: the cost of giving your agent accurate, current, context-efficient information about the world.
02last30days-skill
- Browser cookies are already an API key — tools like last30days-skill use your existing logins to query X, Reddit, and YouTube without paying for API access or dealing with rate limits.
03headroom
- headroom's 60-95% token reduction comes from three techniques: JSON deduplication, AST-aware code stripping, and a retrieval mode that compresses aggressively but keeps the original on hand if the model needs it.
04open-notebook
- Self-hosting NotebookLM via open-notebook unlocks local models, custom speaker voices, and full data privacy — the setup cost is real, but so is the ceiling lift.
08tolaria
- File-system-first note tools like Tolaria let AI coding agents read and edit your notes directly — no API, no sync layer, no app in the middle.
09rilable
- Claude Fable's raw capability is most visible in clone tasks — replicating a full mobile app in two prompts is a useful benchmark for how much less scaffolding modern agents need.
10pm-skills
- Large PM prompt libraries are best used as steal-and-customize starting points — the skill files matter more than the organizational structure.
11openai/plugins
- The OpenAI plugins repo is designed for spec-to-implementation workflows — give Codex a Notion spec and it writes the code.
12Agent-Reach
- Agent-Reach and last30days-skill solve different scopes: last30days delivers a synthesized brief, Agent-Reach gives you the raw plumbing to query any platform inside your own workflow.
14GitHub Star Controversies
- GitHub star counts are increasingly unreliable as a quality signal; PhantomStars detects bulk-starring campaigns, and several popular repos covered this week had suspicious engagement patterns.
15taste-skill
- Design taste prompts replace one set of AI cliches with another — their real value is as a structured starting point for defining your own design system, not as a source of originality.
17microsoft/markitdown
- Converting PDFs to markdown before passing them to an LLM cuts token usage by up to 70%; markitdown automates this and includes an MCP server that does it silently for every upload.
Terms worth knowing.
- last30days-skill
- A Claude Code skill that queries multiple social platforms in parallel and synthesizes what people actually engaged with in the past 30 days, using browser cookies for auth instead of API keys.
- headroom
- An open-source token compression library that strips redundancy from JSON, logs, RAG chunks, and code before they reach the LLM context window, claiming 60-95% reduction.
- AST-aware compression
- Reducing source code size by parsing its abstract syntax tree and stripping non-essential elements, preserving meaning while shrinking token count.
- CCR (Compressed Context Retrieval)
- A headroom technique that compresses context aggressively but stores the original, allowing the model to request full details back on demand.
- open-notebook
- A self-hosted, open-source alternative to Google NotebookLM with support for local LLMs, multiple TTS speaker profiles, and no vendor lock-in.
- Tolaria
- A file-system-first offline markdown knowledge base with a Notion-like UI, designed so AI coding tools can read and write notes directly without going through an app layer.
- rilable
- A proof-of-concept by Riley Brown showing Claude Fable (Mythos) can clone a full mobile app UI from screenshots in two prompts; the name is a portmanteau of Riley + Lovable.
- Taste Skill
- A prompt library that steers AI coding tools away from default generic aesthetics toward named design systems (brutalist, minimalist, soft), with customizable style files.
- apple/container
- Apple's official tool for running Linux containers as lightweight VMs on Apple Silicon, with deep macOS integration for file system syncing — Apple's answer to Windows WSL.
- markitdown
- A Microsoft open-source tool (110k+ stars) that converts PDFs, Office docs, images, audio, and YouTube videos into clean markdown, reducing LLM token usage by up to 70%.
- PhantomStars
- A GitHub bot that flags repositories showing signs of fake engagement by identifying accounts with suspicious single-purpose starring behavior and no other activity.
- Agent-Reach
- A CLI tool that gives AI agents direct read/search access to Twitter, Reddit, YouTube, GitHub, and other platforms at zero API cost using cookie-based auth.
Things they pointed at.
Lines you could clip.
01:23
“Do you want your own little mini Gartner industry analyst on your side? This is that kind of skill.”
04:42
“It is because of Fable. Everyone is seeing Fable is really powerful, but it just burns through tokens. A lot of people have burned through tokens within a day.”
14:24
“As soon as I saw this, my eyes were open to what I could build.”
26:12
“You are really just trading one set of cliches for another.”
30:20
“Each PDF page can consume between 1,500 to 3,000 tokens. So a 20-page document burns through up to 70,000 tokens in one shot, and this is before you even ask your first question.”
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
00:00You know how everyone's got all these markdown files and you can't keep them in Notion, you don't know where to put them, Obsidian gets too complicated, I've got a beautiful free solution that will work for your LLM and that you will enjoy using too. Again, free. Everyone else's website is starting to look like slop.
00:13I've got something that will add taste to your site. And finally, what about when you have your agent try to look up information on the Internet and gets blocked, blocked, blocked, blocked? I've got two free solutions that will unblock it and pull the most important data to you.
00:25That and so much more coming up. Presented by Zapier, the AI automation company.
00:30Peter, the audience loved you last week. It was the number one GitHub repo show that I ever did, so I'm glad that you're back. Thank you.
00:38Let's get into the first repo of the week, which is last thirty day skill. I've been really eager to talk about this. Here's what this is about.
00:44You and I know that when we're doing research, especially for stuff that I'm doing over here, the most relevant information is on sites that are hard to access, like X, which will keep blocking my bots. And it's not just on these sites, it's also within the last thirty days.
00:58Don't give me a tweet that happened six six years ago, or even frankly, uh, six months ago. It's not relevant anymore. What he did, Matt, was he created a skill set right here that makes it easy to go and pull that data, and that's why people are loving it.
01:13I've actually, I think, plugged it in the in the past, but it never really made it into the top 10. It is now number one, the most popular repo of the week. What do you think of this?
01:23I mean, do you want your own little kind of like mini Gartner industry analyst on your side? This is that kind of skill.
01:30You know, it's gonna go around and pull stuff from all sorts of places. Mhmm. You might be like, hang on.
01:35How's it gonna grab stuff from x when I haven't got an API key? I'm not paying for the access. And that's because hidden within this repo, there's lots of little tools for accessing things like x and YouTube and stuff like that.
01:47And it uses your own browser's cookies. So it kind of acts as if it's you Mhmm. Which, you know, is something you might wanna be careful with.
01:56But, yeah, if it works, it works. And it is a it is a common technique to get around the platforms locking you down.
02:02Use the things you've already got that you're logged in with, and then these tools synthesize all that information it finds into a deliverable that then you can read and enjoy and use however you like. I'm pretty sure, though, they did ask me for my YouTube
02:14API key, and maybe even my x. I think it's doing it's doing both, I thought. I could be wrong.
02:21But you're saying I thought part of this is it's not supposed to. Right?
02:24It's not supposed to, but you can bring your own keys into this app. So if you've got keys, then it will use them. But I believe it has fallbacks so that it will use you know, it will try and slurp things out of your browser if you are logged into those sites.
02:37Okay. I I like this, um,
02:39video here that shows Matt actually using it. So Let's do. This is one
02:44that I did right before we started. So I said last thirty days, highest performing cold email frameworks for ICP out output three email variant subject line.
02:56So I went on x, went on Twitter, found Reddit threads, x posts, web pages.
03:02And what's what's interesting how I use this tool is I often don't even read what it says. Like, sure, it's interesting to see what it learned, but mostly I just wanted to write a good email.
03:12So I said, can you write me some cool emails for getting on Greg Eisenberg's podcast? Sorry. I spelled your name wrong.
03:18Good target. Startup ideas later. It's all about unconventional startup ideas.
03:21Community building and fast unique relevance. Now what's your angle? What's your credibility signal?
03:25Any connection points to Greg? Mutual follows. What timely?
03:29Uh, talk about AI tools I'm working on, and I once made a smart oven. Okay.
03:36And then what he's using is he's looking it up, seeing what's hot right now and what's effective for cold email, and then he's having his AI write it for him. Good use case. He's got a few others, and of course, we've got the report here with the full video for people who wanna go see it.
03:50I love that. Next, number two is headroom.
03:54This has been especially popular this week. How would you explain what this does?
04:01So we had this one last week, but it is back. Um, and the whole way that it works is that it uses various forms of analysis of, basically, big things that are going over the wire.
04:14So things like log files and JSON files and things like that, it kind of has these techniques to boil them down into kind of, like, minimal representations of what the agent actually requires to do its job. And then that supposedly will significantly reduce the number of tokens going over the wire.
04:29And there is some, yeah, the truth to this technique. Some of their own benchmarks showed minor gains, but then I believe you have some benchmarks from them that show significantly bigger gains when performing specific tasks like debugging and so on and so forth.
04:46Yeah. It's and I think it's because of Fable. Everyone's now seeing the fable Fable's really powerful, but at the same time, it just burns through tokens.
04:53A lot of people have burned through tokens within a day, all of their allotments, and so this is becoming especially hot. There is an old post here that I'll include in the show notes for people that that just shows that actually, I I asked you, did you understand what he said?
05:14And you said, oh, yeah. What what is he saying here? He's evaluating lots of different tools like this.
05:20So this is a list of various different tools and technologies being used
05:25to, basically, as I said, take things like JSON. So where it says there, SmartCrusher for JSON, DDUP, and structural compression, it's because things like JSON files and JSON documents contain often lots of things that repeat over and over and over.
05:39So if you can strip all of that out, deduplicate it, that's what dedup means, uh, then you can significantly make it smaller, and then smaller equals, you know, fewer tokens.
05:49And it's the same thing with code as well. So where it says there AST aware code compression, that's basically taking the structure of the syntax of code and then boiling it down to just its bare essentials so that you can make the code smaller, again, fewer tokens over the wire.
06:04Okay.
06:05Alright. Let's go on to the next one. I really like this one.
06:11I love NotebookLM because this is Google's project where I can really learn things the way that I like to learn things. My way of using it is I will put a big document into it, and then I end up with a chatty podcast where two people who sound like NPR hosts are talking about and teaching me the topic.
06:27The problem that I have with it is those two people get very boring because you still hear the same two people and the same approach over and over again, and I think that there are other features I would want, and I'd wanna pull it away from Google's framework because I think they're smarter models that I'd at least like to experiment with, and so what they did here was, and here it is, it's open Notebook, they basically said, let's reproduce what Notebook is doing, but give people a lot more flexibility.
06:54I've got a really good video here that I'd like to play a little bit of, and I know people have told me to told you, I guess, or told me that we are we should not be talking over it so much, but I'll let it play a little bit, and when I get annoyed, I'll stop. First, the podcast generator.
07:09NotebookLM made AI podcasts feel actually pretty cool. If you haven't played with it, maybe you should. If I run it here, well, something else happens.
07:18You Take a listen. It's a game changer for researchers looking for autonomy and privacy.
07:23Alex.
07:24I think one of the coolest aspects of OLAMA. Cool. Right?
07:27But Open Notebook gives you more control over that format. You can generate podcasts from your sources, configure the structure, and use multiple speaker profiles instead of being stuck with one fixed stop.
07:42Multiple speaker profiles is is where they got me. Maybe not even have two people. It doesn't have to be two different personalities every time, but even have a third person.
07:51Anyway, it's great. We have a full link here for everybody who wants to go check it out. Any thoughts on this?
07:57I think the biggest win here is privacy. So when you're using Google, everything's, you know, going up to their servers or whatever, but this really pictures itself as being something that you can run locally on local models. Now, you might think, oh, it's gonna not sound very good or not perform very well because, obviously, running local models is very intensive, and a lot of people's machines just can't run them.
08:18Uh, but as we saw, that's getting better. Uh, and last week, we had, I think, something called Vox TTS, which Mhmm. You know, we tried out and produces absolutely amazing voices in such a small amount of space that actually, you know, you could produce voices that are entirely prompted.
08:33You could say, oh, I want it to be someone with a particular accent or, you know, I want it to be yeah. You just describe who it is you want it to sound like. You can't do it with Google.
08:42You can do it with local models. So it is worth a try. You know?
08:47But it is gonna be work setting it all up. So don't think it's just as simple as NotebookLM, know. You've got you've got things to set up here.
08:54That team, by the way, reached out to me, and I love that they did. I'm hoping I get to talk to them. I've got another video here for you to take a look at when when we're done, folks.
09:02And now comes up my comes my sponsor, Zapier MCP. If you are here, you are probably messing around with lots of different tools, lots of different LLMs, lots of different computers even.
09:13In fact, was complaining to Peter quite a bit, as soon as he said hello to me, with how many things get lost. Skills are in this computer. Anyway, the beauty of Zapier MCP is you get all of your connections to all the tools that you love in one place.
09:26It is so easy to use. I really urge you to go and try it out. It's free right now to use it, and you can experience what it's like to connect all your tools and give it the the parameters that you want.
09:36So maybe they don't have write access, maybe they have read access, maybe they have draft access, etcetera, and now you have a URL, a simple tool that you can give to every, uh, every other, um, product that you work with. Go get it. Go check it out.
09:49And thanks to Zapier, we're also gonna get to do these repos worth checking in between here, the top 10. And the first one came from you, Peter. What is this one?
10:00So I first saw this in an X post by, uh, Justine Moore, who is a partner, um, a 16 z. Mhmm. And she was very impressed What what did came out of this.
10:10Yeah. This is the, uh, the tweeter. So what you do is you direct this skill.
10:14It's a codex only one at the moment because it uses codex's built in image generation tools. What you do is you point this to, like, an article, and preferably one that you've written yourself. Mhmm.
10:25And it will go through, and, yes, it's all written in Chinese. I did translate it into English to see what it was about, and it's all pretty straightforward. But you put it at your content.
10:33It works out some metaphors for the different things that you're trying to communicate in that content, and then turns it into these kind of cool illustrations with this black kind of like blob character that you can see, and then it adds annotations. And by default, I think it does actually do it in Chinese, but you can just tell it, do it in English, and it will do it in English.
10:52And I've played with this quite a bit, and I'm actually really impressed with the kind of the metaphors that come out of it. It's, you know, quite impressive. And it's using the GBT image two model for generating the images behind the scenes because that's what Codex uses.
11:05I really like that you found this. I mean, honestly, for me, I love that we're doing the top 10 because it says something, but I also actually, not also, I prefer these great finds that people wouldn't discover otherwise, like this one.
11:18This is called Telaria, and this is another one that you found. Oh, this is so beautiful.
11:22I am so glad you turned me onto this. People are gonna love it. Uh, what's the problem this is solving, and then what is it looking what does it look like?
11:29So lots of people now, me included, have lots of markdown notes.
11:34Uh, you might wanna categorize them and just keep them around, you know, on your system, uh, be able to search them and be able to, you know, look at things and do a whole WYSIWYG markdown editing thing. This does all of that. So there's lots of other tools of this kind of nature out there, but this one is open source and it's free, and it's particularly focused on AI integration if you want it.
11:55I mean, you don't have to use AI with this whatsoever, but you can because all of the files that it uses are literally stored on your file system. So you could have something like, you know, co work or whatever just produce files and edit the files, and it doesn't even have to use the app. Like, it can just edit the files on the file system, then they automatically update in real time within this, you know, note management app.
12:15So very cool. I gave it a run
12:18and really liked it. But, you know, it's a case of migrating over to these things. I've already got my own system, but I would very, you know, much like to try and get that done.
12:25I've seen so many people look for tools that will help them keep all their markdown files organized. Obsidian is the one that's talked about the most, but for most people, Obsidian is too much, and it's it's just not engaging. Meanwhile, look at this.
12:38Pretty, looks simple, looks a lot like Notion, to be honest with you. In fact, I don't know why, but they also have at the top right here under properties Notion ID, so I'm guessing they're pulling documents out of Notion too there. But it is such a good find.
12:52I love that you got that. Oh, there's Luca, the founder.
12:55There he is.
12:57Final one came from me. I'm really excited about this one because as soon as Fable came out, I saw this dude, Riley Brown, push this video on x, which just went insane. Guys,
13:10Claude Fable or Mythos is absolutely insane.
13:15On the left here, we have the actual level lovable mobile app. And here, we have a lovable mobile app that I built in two prompts with Claude Fable five.
13:27This is the chat thread. Here's prompt one, and then I just said run it on simulator two and then go. So that just counts as one prompt.
13:34And then it did some stuff, did some stuff, and then I just screenshotted all the screens on Lovable. I just need you to redesign it to look exactly like this. And
13:45Okay. And it did it. It made it look just like Lovable as you could see on your screen.
13:49But the cool part is that at the end of it, said, you know what? I now want you to build me something. And the thing he said to Lovable to build was Notion.
13:56The thing he said to his copycat of Lovable that Fable built is Notion, and then it went head to head. And you know what? He said, in some ways, I really like mine my version better.
14:06I guess there were some people who said, is this real? He then put it up on GitHub and said, here you go, folks. Go play with it, and you see for yourself if this if this stands up.
14:16It's got 67 stars. I don't necessarily recommend that you replace Lovable with this. Like, get it?
14:21It's Riley able, and Riley's Lovable. But it gives you a sense of what can be done with Fable, and as soon as I saw this, my eyes were open to what I could build.
14:29I love this. What do you think?
14:31I don't have a lot to add to that, really. I mean, this is clearly one developer's passion project, and, you know, that's one of the reasons that people stick with things like Lovable and Bolt when they get them working is because, you know, they know there's a massive amount of money on the table and there's a big team and all that type of thing.
14:45Right. But it is great to see that just individual developers can spin, you know, spin things like this up It's totally within your power if you've, you know, got the skills to do it. Don't use it to replace Lovable.
14:55Use it to see what you can Not just yet. No. Okay.
14:57So now, number four, repo of the week is PM skills, 68 skills, and 42 chain workflows across
15:05nine plugins. I was like, ho about it, and I was a little sheepish, and then you said, no, there's a real value here. What do you see that's valuable in this?
15:15Again, you know, this is another one of these repos that's full of different skills and prompts. Uh, it's basically a big prompt library. But this is clearly built by someone who, you know, has a little bit of knowledge about project management.
15:28I know nothing about project management, so any guidance that I could get is valuable. Um, but it doesn't it does have some commands in it, so you can do things like, uh, get it to interview you about your project. And so it will just keep drilling you with questions until it can put together a a plan, um, and it will do things, um, you know, like help you do consulting and strategy and all those types of things.
15:49And I don't know those areas too well, so this is useful for me. But I would say if you do have any knowledge of those areas, this is another one of those repos to go through, read, and steal from because you might not wanna take on their approach to project management wholesale.
16:05You might wanna come up with your own thing, and this will give you a good basis upon which to build your own, you know, tooling to do those interviews on yourself.
16:13Well, so, you know, I usually like to find YouTube videos that illustrate this to see how other people use it. I found one. Guys went through it, and it was terrible.
16:22And then I got in my own head, I go, I'll take it all back. No. No.
16:25No. No. It wasn't the skill that's terrible.
16:27Their video was skill was terrible. So I go, oh, no. They've they've got a lot of views, but it makes no sense what they're saying.
16:33They're just, am I that bad? And then I started getting in my head about, like, I've gotta get better here. I've gotta prepare more.
16:39And then I drove myself and my family crazy this morning trying to, like, go through this and not be bad and not so, anyway, I've got no video on this. All I got is a head case on this, and I'm working through it.
16:50So let's go on to the next one. OpenAI plugins. This is another one that I didn't fully comprehend the value of until you and I talked, and it's from OpenAI.
17:00It's plugins for their software for Codecs. What is this?
17:06Well, of course, I just wanna firstly say that, you know, everyone should go and use your sponsor Zapier. But this does provide some similar skills if you're willing to kind of get, you know, deep down with Codecs, and you wanna plug in things like Asana and Airtable, like you can see on the screen there, but also things like Gmail and g Calendar and different research things as well I noticed were in there, like, things I would never use, but there's, like, research APIs and things.
17:32If you wanna provide access to those, to Codecs, these are the plugins to do so because, you know, they're kind of semi official or perhaps even fully official. I don't really know what the third parties have to say about it, but this is what OpenAI is putting forward.
17:45They're saying, like, you know, you wanna hook into, like, Notion, for example, then install this plug into Codex, and off you go. You can ask it to do this and do that. So, um, you know, this is kind of almost like a a homebrew local version of some of, like, what tools like Zapier do in connecting things together.
18:02But,
18:03yeah, you kind of have to figure out all the moving parts for this. What I also saw was that it does it's it's about getting data out of these tools to do something with the tools. It's not just open ended.
18:15So, for example, I might connect Notion directly into Claude, and that's fine, but what they're trying to say here is if you wanna turn a spec into implementation in Claude, we also are gonna have that covered, and this will find the specs, turn it, and start to implement.
18:32That's it's like a not connect it's not about the connection from what I understand. It's about the utility with guidance.
18:39Am I getting that right?
18:41No. I totally agree. I mean and this is the type of thing where you can use these tools within your other workflows that you're doing.
18:48So let's say, you know, you're building, uh, an app in Codex, and you want it to use your project management system, like, say, Jira or something like that for doing tickets. And so, you know, you could you can tell it, you know, go and look at the tickets that are currently on our system Mhmm.
19:04Pull them down, implement something, and then update the ticket with comments or whatever it is that you wanna do. So it just allows you to plug in these third party systems into your Codecs workflow.
19:14Okay. Alright. There's our launch video.
19:17I'll leave it for everyone to get in the show notes. We're gonna move on to agent reach. This is one CLI that lets your agent read and search Twitter, Reddit, YouTube, GitHub, and other sites.
19:29You you've this you made a distinction between this and the first 30 in the last thirty days skill sorry, the last thirty days repo that we talked about in the beginning. What's the difference here?
19:40When I looked through the code, I just kind of saw some similarities between the two repos. So that last thirty days thing, you basically give it a topic, and it produces a deliverable Mhmm.
19:50Of, you know, what happened in the last thirty days in whatever topic you want. But underneath the hood, it's using a bunch of tools to access things like x and YouTube and so on, and pulling data from those to synthesize that report. Well, this is like if you just took that layer of tools and just turned that into a thing.
20:07So if you want more granular access, if you wanna be able to ask Claude or whatever, you know, go and look on x for such and such or look on YouTube for whatever it is that you want, then this is the kind of the more surgical like, the surgical knife that allows you to do that without producing that big deliverable.
20:25You just use it within whatever workflow you're working on at this current moment.
20:29I like this example for LinkedIn. Tell your agent, help me set up LinkedIn. Yeah.
20:35I again, this is for researching, for understanding somebody.
20:39When I talk with Adam about tools like this and I say, how do you use it? He says that they use it all the time in the venture studio because one of their companies is trying to hire, so they wanna research people. One of their company is trying to buy a company or work with a company, and so they need good research tools.
20:53They don't wanna have somebody clicking a mouse on a screen, and that's where this comes in. Make sense? And it does a few yeah.
20:58There's a few Chinese social networks as well. I don't know anything about these networks. Mhmm.
21:02But I believe the developer is Chinese, and so they've added in these various other sites that I'm completely unaware with. Like, what's it? Billy
21:10Billy Billy or something? I don't know. I know nothing about these But guess, I you know, if you want access to them, then great.
21:17Yeah. I don't know if you noticed, but when I read off the list of sites that you can read from, I intentionally said and others. Once we get to the Chinese ones to make sure that I didn't mispronounce them.
21:24I'm not familiar with these. Okay. The tool still works.
21:27Really popular, and it's number six this week. Again, download everything in the description. Let's go on to this one, less applicable to our people, but still worth knowing about.
21:36This is NVIDIA's open platform for world models, data sets, and tools for physical AI. We're now talking about robots, autonomous vehicles, smart infrastructure.
21:45I've got kind of a video here to to show. Anything you wanna say about this?
21:51Not really, other than the fact that this is a new, of what they call an omni model. So it can handle lots of different types of input, not just text, but also, you know, like, within space and images and all this type of thing.
22:04So it's good for people that are working in the physical space, which is not me. I know absolutely nothing about robotics.
22:10But you wanna look at robot arms moving around, then this is the place to do it. Yeah.
22:14They do have that on their site. And I do wonder They do. People who are who are listening are doing that.
22:19I'll tell you this, folks. I stopped talking over the videos because one commenter told told us that and Peter Peter noticed it and told me, if this is something that you're building with, let us know.
22:30Let us know how you're using this. Of NVIDIA Omnidreams,
22:33an action conditioned world model. Cosmos predicts the future frame by frame.
22:40Post train Cosmos, and it becomes a world action model. Perceiving, reason, planning, generating actions.
22:49Okay.
22:52Um, I like that it's here. I like that we're getting more substantive and getting offline. I don't think it's applicable to most people.
22:58We'll move on to number eight. Oh, before we do. We keep commenting on how some repos don't seem to fit here.
23:06How do they get all these stars? And you have said, look, some people do Adam actually said some people buy. You've been questioning some.
23:14And then you you highlighted these. What are we looking at here?
23:19So I noticed that this Phantom Stars account, which I know nothing about what this is, but it's a tool of some kind that is going around and commenting on different repos where it thinks there has been what it calls fake engagement. So it's not making a direct, you know, accusation, and we're not making a direct accusation that these projects are engaging in any kind of, like, buying styles or whatever.
23:41But it's saying we'd noticed that there are these users that are very suspicious and do nothing but go around and, like, random repos and have no other activity. So I just thought it was interesting that it cropped up on a few of the ones that we've covered today. But I guess that makes sense since they are the most popular ones of the week.
23:57They're the ones that are most likely to attract this, you know, whether or not it's being done intentionally or not.
24:02Yeah. They're gonna get the most everythings. Okay.
24:05So we've got links here for those. It is happening. It's a thing.
24:08Yeah. And if anyone has any information about this, let us know. I'm curious about who's buying stars.
24:13What does stars cost? I've been hearing a lot about people who are buying engagement on x, and it doesn't cost that much.
24:19Some anywhere from 50 repo for a thousand dollars. So, you know, if if anyone wants that help, you know They want bots to do it. They want a thousand repos started for $1.
24:29Okay. This is bringing up TasteSkill is bringing up a really big issue, which is that everybody's design looks the same, especially if it comes from any of Claude's products, except I don't know about Fable.
24:40I've been hearing Fable's getting better at it. But if you wanna create your unique taste, if you wanna create your unique angle, don't start with what Claude has and adjust it. Start with something brand new.
24:49That's the vision here. And they've got here, let me show you a before and after from this video.
24:56So you can see this is a very basic landing page. So even though I think it's decent, but it's pretty basic, so let's check out the one that has the test scale.
25:14Taking a moment, but he'll get it.
25:16And you can see that this is even better. So this is actually really premium, and you can see that the UI is very elevated.
25:24So it's very good. So
25:28And that's that's the idea here. You can take a look at their site, which, of course, we'll link to, and and you can get a sense of their design style. I like it.
25:36I like these a lot. What do you think?
25:40I think it's very important to note that you are basically trading one set of cliches for another with skills like this. So, like, when I went into and looked through the skill files, you know, it's got all these different kind of, like, tones, like brutalist design and minimalist design and a soft design, which you can see on that page as well.
25:57Mhmm. And if you go into the skill files, it basically says, oh, you know, like a soft design will use these types of colors, and it will use these types of fonts. But you're really just over you know, steering the model away from what it would do by default into this other kind of lane.
26:10So you're just . But the good thing about it is you can take repos like this. You can go in and say, well, actually, I wanna make my own style inspired by all of this, copy and paste it, change the fonts around, change colors around, and then bam, you've got your own little design system come out of it.
26:26So I don't mind repos like this. It is nice to see what other people's takes are on designing things. And it's kind of like fun to play with.
26:34You know, it's not like it's producing code that's gonna completely break everything. It is basically just changing the design, the style, the CSS, the fonts, so on. So, yeah, experiment play.
26:44But also, again, you know, I've said this on so many repos, it's a case of go into the repos, look at the skills, and see if you're happy with what it's suggesting, because you might not like you might not like these designs, but you can change it.
26:55What I like about this video from Devs Kingdom is he he shows how easy it is. He just has it create another version, he clicks it open, and he shows us what it looks like. And I think that's another fun thing about this.
27:06Just throw your design at it, see what it can do, and then adjust. Alright. This is Apple's own tool to run Linux containers as lightweight virtual machines on Apple silicon.
27:20This came out at WWDC this week. How and what's going on with this?
27:26So, I mean, this idea, um, even at Apple, has been actually around for, you know, a year or two now, which is that it allows you on macOS to basically run Linux within a tiny virtual machine within your machine.
27:43Now you've been able to do this with third party tools for a long, long, long time. This is, you know, a technology that goes back decades. But now Apple's getting in on the action, and people are thinking, well, why?
27:52Why would they wanna do that? And that's why the announcements at WWDC this week in particular were interesting is because they're taking it to the next level.
28:01So it's not just about running, you know, a random Linux in a VM. You've out to do that for years. They're focusing more on the integration now between macOS and these underlying VMs.
28:12So they're adding in things like being able to keep the file system synced between the two. And hopefully, you know, they will eventually get it to the point where it's a bit like there's something in Windows called WSL, which allows you to run Linux, and it fully integrates with Windows.
28:27You can even run graphical apps from Linux, and they appear on the screen, you know, almost as if they're Windows applications. The idea is that, hopefully, you'll be able to do all of this as well with macOS eventually. They're just kind of, like, teasing out bit after bit.
28:41And Tim Sniff there is one of the people that's actually working on this feature. I mean, one of the things that actually I found quite funny about this tweet is that he said, oh, you know, I love what's going on with Apple container. And someone replied back and said, oh, you know, we've been able to do this for years, And people didn't realize he's the guy actually implementing this stuff.
28:57So, um, it was one of those kind of scenarios. But, uh, it's nice to see that Apple's interested in this, and they're allowing their people to talk about it. That's quite unusual, the Apple with technical stuff.
29:08You you don't often see them tweeting about things that they're actually implementing. So it seems like they wanna engage with our community, and they want, you know, developers and particularly AI developers to use these containers. I mean, they're technically VMs, um, to run workloads safely within a macOS machine without it taking over the entire machine, and that is one of the things that this allows you to do.
29:29Alright. Repeat. Mark it down.
29:33Coming back again. This one keeps making it into the top 10. This is Microsoft's way of helping you turn PDFs into Markdown, Office Docs, images, audio into clean Markdown files, even YouTube videos into Markdown files.
29:49We've we've talked about it before, so you know what? Today what we're gonna do is we're gonna have this YouTuber who I I can't tell if this is a real guy or not because he's so like he's so always on exactly right here, so maybe he's super edited.
30:02Everyone uploading PDFs to Claude is basically wasting their tokens, and Microsoft had a free solution to this for months. So every time you upload a PDF, Claude has to process the entire thing, the formatting, the broken tables, the images, and all the extra junk inside the file. And each PDF page can consume between 1,500 to 3,000 tokens.
30:21So 20 page document burns through up to 70,000 tokens in one shot, and this is before you even ask your first question. The fix is simple. It's called Markitdown, a free Microsoft tool with over 110,000 stars on GitHub.
30:33It takes any file, be it PDFs, Word docs, Excel sheets, PowerPoints, or even YouTube videos, and converts them into clean markdown text. This drops your token usage by up to 70%. And Claude actually gives better answers because it was trained on millions of markdown documents and understands the format natively.
30:50Plus, markdown comes with an MCP server. Meaning, when you connect it to cloud desktop, it automatically converts every file you upload to a dot m d file. So check the pinned comment for the tool link and the full PDF guide.
31:03Alright. I like his his way of doing it. Do you think he's human?
31:07I'm not sure, but are you human?
31:09I would like to be I less definitely am. Yeah. I would like to be less human.
31:13I'd like to get like more AI into the scripting here. Look at how organized he was with that. Alright.
31:18Of course, that report with his video and everything else is gonna be in the description. Before we go, Peter, what are you up to this weekend? I am literally doing nothing.
31:28I'm attending a family barbecue, but other than that, I'm doing nothing, because I just need to decompress all these stars are literally getting into my head. I need to, like, get rid of the stars. Love that.
31:37Yeah. I am going to be mowing, mowing, mowing this weekend, doing a lot of yard work, and then Olivia is taking the two older boys with her to San Francisco to be with her family, and I've got the baby with me for ten days, and I don't I've get how we're gonna done that before. Yeah.
31:53It's it's something. But, you know, he's in daycare, and then I think I think on the weekends, I'm gonna take him to San Antonio.
32:00We're in Austin, and I'll take him to a museum where he can run around all the time, and I'll bring my guitar with my earphone so I could practice and just have some downtime while he's running. Alright. So there it is.
32:10As the next video, now that this is over, you should know that there have been some incredible AI releases this week that did not get much attention because Fable was big and also because people just don't have that much attention span. We found the best ones in the next video that you're gonna watch. You're gonna see them.
32:24See you in there.
The Hook
The bait, then the rug-pull.
Every week, the open-source community quietly ships tools that would cost a hundred dollars a month if someone put them on a Stripe checkout. This week's haul includes a NotebookLM you can run on your own hardware, a single CLI that searches X and Reddit without an API key, and a compression layer that cut one developer's token bill by 95%.
How they asked for the click.
VERBAL ASK
09:00product
“Go get it. Go check it out. And thanks to Zapier, we are also gonna get to do these repos worth checking in between here.”
Mid-roll dedicated chapter. Andrew describes the product personally, mentions it is free to try, frames it as solving his own problem of tool fragmentation across machines.
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.
