Modern Creator
The Next New Thing · YouTube

Cut LLM cost by 95%, replace ElevenLabs, and 10 top GitHub repos

Andrew Warner and Peter Cooper rank the week's top 10 AI GitHub repos and debunk most of the headlines.

Posted
yesterday
Duration
Format
Interview
educational
Views
3.4K
175 likes
Big Idea

The argument in one line.

The most trustworthy AI GitHub repos are boring, extensible tools from established teams -- headline numbers like '95% cost savings' almost always collapse to 4-8% in practice, and anonymous skill bundles with 200K stars are more dangerous than useful.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…
  • You build AI agents or use Claude Code daily and want a curated weekly filter on what is actually worth installing vs. what is hype.
  • You pay for ElevenLabs and want to know whether free local voice-cloning has caught up.
  • You are frustrated that every AI-generated UI looks identical to Anthropic's default design system.
  • You manage dozens of clients or contacts and want a structured memory layer for your AI agent.
SKIP IF…
  • You want step-by-step technical tutorials -- this is a conversational roundup with screenshares, not a build-along.
  • You are new to AI tools; the show assumes working familiarity with Claude Code, MCP, and GitHub.
TL;DR

The full version, fast.

Two hosts rank the week's top 10 AI GitHub repos and debunk inflated claims throughout. Headroom, promising 60-95% token savings, delivers a median 4.8% in its own benchmark of 50,000 sessions. The genuinely useful tools are Microsoft's markitdown, PewDiePie's self-hosted AI harness Odysseus, the local PDF parser liteparse, design taste files for breaking Claude's aesthetic monoculture, and VoxCPM for free local voice cloning. The co-host's consistent warning: read every skill-bundle rule before installing -- someone else's rules will silently override your voice, your coding style, and your workflow.

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 →
Chapters

Where the time goes.

00:0000:36

01 · Cold open

Teaser clips for MoneyPrinterTurbo, Odysseus, and VoxCPM set up the episode's three headline items.

00:3602:42

02 · #1 MoneyPrinterTurbo

Type one sentence and the tool writes a script, finds footage, adds voiceover and subtitles. Functional but produces AI slop. 700MB install; needs API keys.

02:4204:39

03 · #2 headroom

Token-compression proxy claiming 60-95% savings. Their own 50,000-session benchmark shows median 4.8%.

04:3907:03

04 · #3 markitdown (Microsoft)

Converts any document to clean markdown. Extensible plugin system; Azure APIs by default.

07:0308:18

05 · Sponsor: Zapier MCP

Zapier as a trusted middle layer giving agents access to 8,000+ tools with user-controlled restrictions.

08:1811:51

06 · Honorable: Odysseus (PewDiePie)

Self-hosted AI harness with memory, deep research, gallery editing, web access. Too technical for average users but genuinely impressive.

11:5113:57

07 · Honorable: gbrain (Gary Tan)

Structured knowledge graph for AI agent memory. Postgres + embeddings + BM25.

13:5715:54

08 · Honorable: Webwright (Microsoft)

Browser agent framework using DOM code instead of pixel screenshots -- more token-efficient.

15:5417:06

09 · Honorable: liteparse

Fast, local-only PDF parser from LlamaIndex. No cloud, no API key, handles complex layouts.

17:0619:21

10 · #4 compound-engineering-plugin (Every)

63 agents, 249 skills. Trusted source but Peter's rule: read the skills yourself before installing.

19:2121:57

11 · #5 stop-slop

Five text files stripping AI writing tells. 7,500 stars. Peter warns some rules are actively bad.

21:5723:57

12 · #6 supermemory

Long-term AI memory across every tool. Peter prefers manual context injection.

23:5725:57

13 · #7 ECC

63 agents, 249 skills, ~205K stars. Peter: too much, too unknown, no trust.

25:5729:33

14 · #8 taste-skill

Design skill files to break Anthropic default aesthetics. Good for inspiration; build on top.

29:3332:06

15 · #9 Understand-Anything

Codebase to interactive knowledge graph. Hacker News skepticism about star count authenticity.

32:0634:57

16 · #10 VoxCPM

Open-source local voice cloning. 30 languages. ElevenLabs moat is compliance paper trail, not quality.

34:5735:23

17 · Wrap-up

Full report link in description, Peter's handles, plug for last week's episode.

Atomic Insights

Lines worth screenshotting.

  • Headroom's own benchmark of 50,000 sessions shows a median 4.8% token saving -- the 60-95% headline applies only to log-heavy debugging sessions.
  • Installing a skill bundle without reading it hands someone else's opinions control over how your AI codes, writes, and thinks for you.
  • Microsoft's markitdown treats document conversion as a pluggable framework problem, not a one-tool job -- and that extensibility is the real value.
  • ElevenLabs' moat is not audio quality; it is the legal compliance paper trail proving all voices were licensed.
  • VoxCPM is open-source, runs locally on a consumer GPU, produces studio-quality audio in 30 languages, and requires no API key -- ElevenLabs charges $22-$990/month for equivalent output.
  • PewDiePie built Odysseus because local LLMs lacked memory, deep research, and decent UX -- and actually shipped something useful.
  • GitHub star counts are a broken signal: repos can buy stars, and 200K stars on an unknown name should raise suspicion, not trust.
  • LLMs extract insights from existing artifacts better than they generate new ones -- tools like Understand-Anything lean into that strength rather than fight it.
  • The right way to use someone else's AI skill file is to read every rule, take the ones that fit, and discard the rest -- not to install the whole bundle blindly.
  • Structured memory tools are most useful when managing dozens of people; for focused solo work, manual context injection keeps sessions cleaner.
  • Automatic AI memory can backfire: the model uses past context to reframe current questions, which breaks independent queries.
  • Markitdown's killer feature is turning a scanned image inside a PDF into a text description on the fly -- something most document parsers skip entirely.
Takeaway

Twelve repos, one honest filter.

WHAT TO LEARN

A headline number from a GitHub README is almost never the number you will see in your actual session -- and the repos worth installing are the ones whose every rule you have personally read.

02#1 MoneyPrinterTurbo
  • A tool that generates TikTok-ready videos from a single sentence is functional, but the output is recognizable AI slop -- whether that matters depends on whether your audience cares about authenticity.
03#2 headroom
  • Headroom's own benchmark of 50,000 real sessions shows a median 4.8% token saving -- the 60-95% figure applies only to log-heavy debugging sessions with bloated inputs.
  • A compression proxy between your agent and the LLM only saves money when traffic is genuinely bloated; in ordinary coding sessions, the overhead can cancel out the token savings.
04#3 markitdown (Microsoft)
  • Microsoft's markitdown is worth using because it is a pluggable framework, not a one-trick converter -- swap in your own processing logic for any file type the default misses.
06Honorable: Odysseus
  • PewDiePie built Odysseus because the UX of local LLMs was broken: no memory, no deep research, no web access. The result proves local AI is ready for technical users.
07Honorable: gbrain
  • Structured AI memory is most valuable when managing many contacts or projects; for focused solo work, manual context injection keeps sessions cleaner.
10#4 compound-engineering-plugin
  • Before installing any AI skill file, read every rule: someone else's rules can silently ban adverbs, rewrite WH-sentences, or impose coding conventions you disagree with.
11#5 stop-slop
  • A skill file that removes AI writing tells is only as good as its rules -- some in stop-slop are counterproductive, and any such file needs updating as models change.
12#6 supermemory
  • Automatic AI memory can backfire: the model uses past context to reframe current questions, which breaks independent queries. Manual context injection keeps control with the user.
13#7 ECC
  • A bundle with 200K stars from an unknown author is not more trustworthy than 500 stars from a team you know -- star counts can be purchased.
  • By the time you review a 249-skill bundle thoroughly enough to trust it, you have learned enough to write your own version.
14#8 taste-skill
  • Study taste-skill outputs, take what fits your aesthetic, and build a custom file from those observations -- do not install the whole bundle.
15#9 Understand-Anything
  • LLMs extract insights from existing artifacts better than they generate new ones -- tools like Understand-Anything lean into that strength rather than paper over it.
16#10 VoxCPM
  • VoxCPM runs locally, requires no API key, and produces studio-quality audio in 30 languages -- ElevenLabs real advantage is the legal compliance paper trail, not the quality gap.
Glossary

Terms worth knowing.

MCP (Model Context Protocol)
A standard protocol that lets AI agents call external tools and services. A single MCP URL can expose thousands of integrations the agent can invoke during a session.
Token compression proxy
A middleware layer between an AI agent and the LLM provider that compresses or summarizes input before it is sent, reducing token count and therefore cost.
Skill file / skills bundle
A markdown or structured text file that instructs an AI coding assistant how to behave -- covering code style, commit format, writing rules, and domain knowledge. Can be shared as a GitHub repo.
BM25
A classical text-ranking algorithm used in search engines to score document relevance based on term frequency and document length. Often combined with vector embeddings for better retrieval.
RAG (Retrieval-Augmented Generation)
A pattern where an AI model fetches relevant documents from a database at query time and includes them in the prompt context, rather than relying solely on trained knowledge.
DOM tree
The structured code representation of a web page that a browser uses internally. An agent reading the DOM sees the page as a hierarchy of elements, not pixels -- making browser control faster and cheaper than screenshot-based approaches.
Resources

Things they pointed at.

Quotables

Lines you could clip.

03:21
The actual median saving is 4.8%. The headline is making it bigger than it is.
Tight debunk of a viral GitHub claim -- no setup neededTikTok hook↗ Tweet quote
25:04
By the time I have gone through it, I might as well have made my own.
Universal line about over-engineered toolsIG reel cold open↗ Tweet quote
30:06
LLMs are much better at reading things than writing them.
Counterintuitive claim that reframes how to use AI toolsTikTok hook↗ Tweet quote
34:03
It is producing a paper trail -- that is basically what it is.
Reframes ElevenLabs value prop in one sentencenewsletter pull-quote↗ Tweet quote
The Script

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 tool that will make videos for you for free and claims to print money for you. PewDiePie, the famous YouTuber says get rid of chat GPT. He's got a free replacement and it's amazing.
00:10If everything that Claw designs for you just looks like everything Claw designs for everyone else, I've got a taste repo that will fix that and make things look beautiful. And watch out 11 Labs, I have for you a simple way to create voices. Clone your voice or anyone else's and have AI read in that voice.
00:27So much more coming up today. Presented by Zapier, the AI automation company. Alright, Peter.
00:33Top 10 GitHub repos of the week starting with number one of the week is yeah. Money Printer Turbo. I don't know how I feel about this one, but I'll explain what it is.
00:43You go in, you type what you want a video to look like, you want it to sound like, and it will actually create the freaking video for you. And in many ways, it is impressive. But once you get into the money printing thing, it feels like it's overhyped.
00:56I'm gonna show it to you on GitHub so that you get a sense of it looks like. I'll show you a video of how it works. And I've talked about this before, Peter.
01:05This is a way of creating what many people would call AI slop. Maybe I'm being too harsh about it. I don't know.
01:10What do you think about this?
01:12Well, I was hoping that it was just gonna, like, print out money, you know, one of those money guns you get, but, um, unfortunately not. Um, no.
01:20I mean, in terms of, uh, the the overall project, it's actually quite well structured, it does what it says it's gonna do. But the problem is, do people wanna watch these videos? I guess that's the key thing, you know, you as a video producer.
01:31It's like, people wanna watch this stuff, or would they rather watch someone actually do it for real? You are gonna need a bandwidth, though, if you download this repo just because it has it's like 700 megabytes in size.
01:42Discovered I actually cloned it, and it's full of fonts and samples and stuff like that. So it's a big one.
01:48It's fairly easy to use too, though, once you do get it on your system. Right? You just fill out a form or you just yeah.
01:53There it is. You just give it what you need
01:56and it creates it. Okay. Maybe overhype title, but maybe that's how you become number one.
02:00Yep. Or But you need to plug in some API keys as well because it needs to use, uh, OpenAI or Deepsea. It can it can support a whole different bunch of APIs, but, obviously, you are gonna be paying for the the output from this.
02:11But it will also one of the cool things about it, though, is it will pull in, um, I don't know if you'd call it royalty free, but kind of like free video that's on the web in certain repositories, and then it will integrate that into what it produces.
02:24So it does some mildly clever stuff even if it isn't quite actually printing money.
02:30Okay. Alright. Maybe I'm being too harsh on it because it's promising a lot.
02:33But, truthfully, if you can generate even b roll like this and find it, I think there's value to it. Okay. That's number one for the week.
02:40Number two is headroom. Your AI agents are burning through money reading junk, huge logs, bloated files, repeated search results. This sits between your agent and AI company and shrinks everything before it's sent, claiming 60 to 95% off your token bill with the same answers.
02:55Would you try this? What do you think of this?
02:58It's an interesting idea. I like the idea of putting, like, a proxy in between your agent and the LLM, and then doing things with that kind of gap between to make the the communication more efficient.
03:11Just one thing I should pull up on those. In that headline, it says something like reduce your token bill 60 to 95%, which Yes.
03:19If you actually go into that they've got a page on their GitHub repo, which is actually a a real benchmark of what they've actually done with 50,000 different sessions that have run through this app. K.
03:29And the actual median saving is 4.8%. So the headline is making it bigger than it is.
03:36But, mean, I'll just quickly explain what it does just in a technical sense is let's say that your Claude code or whatever it is is sending lots of, um, log files or JSON or something over the wire.
03:48This thing will look at that and then compact it down to a more minimal representation that the LLM actually needs. So for a certain type of debugging session where there's lots of logs flying around, say, it may be able to reduce it by 60 to 95%.
04:02But in general day to day use, um, no. Not so much. Um, and that's by their own numbers.
04:10Would you actually use it? Is there a need for it? I don't think I use Claude Code enough to need something like this.
04:15But what about someone like you?
04:18Personally, I wouldn't trust it, but that's just that's just me being, you know, critical of these types of things.
04:27But, you know, if it works for him, then great. I mean, there's people that would love this type of thing. Um, I just would rather not have something sitting in between me and the LLO.
04:35Right.
04:37Especially for such little value. Alright. Let's go on to number three.
04:40Mark it down. This one I absolutely love. It is Yeah, I like Mac from Microsoft on the top 10 list.
04:46What it does is it takes a PDF, a Word doc, a PowerPoint, Excel, whatever, and it converts it into markdown so that it's easy for you to to work with. It's easy for you to to send to your agent so that you don't end up taking up, uh, so that you don't end up paying for a lot of tokens just to have your agent convert it.
05:04Uh, I've got some examples of it here. You take a look at actually, here. Let's take a look.
05:09Look. Oh, this one I wanted to show. It even took this image on the right, and it turned it into a description on the left.
05:16In this two panel comic, a cartoon dog sits calmly at a table, admits, etcetera, etcetera. I like that it does that. I've got a short video here, uh, that I found on YouTube with someone showing how it works on a PDF doc.
05:28Look at how beautiful this is.
05:30I'm gonna mute it while it works. All in my virtual environment. I've got a PDF here, just a doc PDF, and I can run this in my terminal.
05:38I'm gonna run mark it down doc PDF output m d. That's it. It makes me a file automatically.
05:46We can open that file, and inside is sort of what we're hoping to expect here. Headings are clean. Tables actually look like table.
05:54Just works. What do you think of this one?
05:58Right. The real win about this, other than it being from Microsoft, is it's very extensible. And so what I mean by that is that it's kind of like having one program that you can just throw any kind of, like, media or spreadsheet or Word document or whatever and then get out what the agent wants from the end of it.
06:15The actual value isn't in doing that conversion because, I mean, you can take any LLM and it will convert an image into some text. But this is so that you can just throw stuff at one thing and get stuff out. The only catch with it is that by default, it uses, um, mostly Azure APIs for doing things.
06:32Like, let's say you've got a scanned PDF, it will use an Azure API to do the conversion. Um, however, because it is so extensible, you can write your own plugins for it, so it would use whatever you want.
06:43So if you've got some magical way of converting, you know, an image into text or whatever, you can add your own plugins to this and build on top of it. So it's kinda like a framework. Mhmm.
06:51Um, and I just like it from that perspective. It's not just one tool for one job. It's like one tool that you can just throw data in, get stuff out, but then extend it with your own building blocks.
07:02Peter, I freaking love the way that you explain these things. Alright. Let's go on to the next one.
07:05Oh, the next one is my sponsor. It's actually why I'm in a hotel room today. I'm here to meet with the founder of Zapier to talk about how we've been doing with them, and I'll tell you one of the things that excites me and why I think we've been doing really well.
07:16Peter, my problem is that I keep moving from, like, Claude to Codex to this to that, and I wanna bring all of my tools along with me. I don't use Asana, but I do use Gmail, and I do use Google Calendar, and I give it access, Notion, and so on. I wanna be able to hand off all of these tools to whatever new thing comes out and know that I both give it the power that it needs, but also have the restrictions that I want to to give it.
07:38And that's the beauty of this. The thing that I love about Zapier is they've now become, in addition to all the automations we know them for, they become that trusted middle layer that lets me give permissions to my agents, but also restrictions. They cannot send email.
07:51They cannot delete all my files and so on. Fair?
07:56Fair. And, actually, it kind of ties a bit into that last item that we did in that it's something that sits in the middle, inputs, outputs.
08:04I mean, yeah, there's a common pattern, actually, with a lot of the projects we're covering.
08:09And a dependable company to do it, and not just for me, for people on my team when I wanna protect them because they're not as deep in it as I am. Alright. And because of Zapier, we're gonna do a couple of honorable mentions here in the middle of the top 10.
08:19Number one honorable mention, I don't know why this does didn't make the list. This is PewDiePie. He created Odysseus.
08:25It's a chat GPT alternative that runs on your computer. I think this is really creative. This is the guy who used to play video.
08:32He actually still, I think, does play video games on YouTube. And here's his realization. He said, actually, here.
08:38Let's have let's play him with his realization.
08:42It's crazy. The point is you realize the model that you can run at home are amazing. The problem is the way you interact with the model.
08:49Okay? Because I would at the time, I was trying to self host the UI for it to get the same experience as I was paying my subscriptions for. And I should say this is because he's got these these LLMs that he realized he doesn't have to pay for.
09:02He can put on his computer. He doesn't have to pass his data outside of his own little universe, but it's really bad, uh, user experience.
09:09And I realize all these things that sort of come together and make, uh, this an amazing experience was completely missing. Where's my memory? Where's my deep research?
09:19What about the agent? What about all these simple integrations like Weblox? I got spent, like, two days trying to integrate that.
09:26So that's why I started to build my own. And I thought it was funny as a meme. I'm like, look.
09:31That could be out there. It's building that. And
09:35then he did it, and it's actually really good. Just unfortunately too techy. Let me show people just a quick video of what it looks like on the inside.
09:44And I'm going to add a message.
09:46Right. This looks a lot like the Claude code experience. So check this out.
09:49Requesting the request.
09:51Here's gonna give me a response in a second. Should say this is, the Claude chat I'm using.
09:56I can see things like amount of context that's used. I can regenerate from here, rewrite shorter, or explain simpler.
10:03Let's see how that works. Okay. So it's an in place
10:08I know this looks basic, but let me show you this for a quick review. Which is nice. We can edit.
10:13Oh, interesting. We can edit the AI's response. I like editing the AI response.
10:18I was talking to Eric Reese who's got, um, an AI project that also asks that also enables you to edit the response, and he said it's because everything that is in a chat experience is context for the next response that you get. And he goes, if you just let it have b s in there, it then uses it to understand, and so you wanna edit.
10:36But that's a small feature that I liked. I will actually zoom in and show you on the left. It does way too much.
10:41It will do notes and tasks. It actually has where is it?
10:46It will do gallery which then lets you edit photos. It's just it's very geeky in the most beautiful way.
10:53It's perfect for our audience. I don't think it's ready for the average user. What do you think of this?
10:59It's an increasingly common thing to do. I mean, it's almost becoming like the the the build your own blog system of, like, this decade. It's like, oh, build your own agent, build your own harness, build your own thing.
11:10And it's good to see someone that is as opinionated as PewDiePie doing it. I mean, I wouldn't necessarily use it because I don't necessarily have the same, you know, values or approach that he does, but I can see why he's doing it. And so it's very exciting to see, because he's someone, you know I think he's still one of the most subscribed channels on YouTube.
11:28Yes. Hundreds of millions of subscribers. So it's just fun to see what he does with that freedom and, you know, the wealth that he's got.
11:34And the fact that he's messing around in what kind of we're doing is, uh, really fun.
11:39Yeah. And, you know, if you wanna do if you do want privacy and lately, I think about all the medical stuff I send into Claude, I get it. True.
11:48Yeah. It's a good good tool.
11:49Alright. Next, g Brain. Again, honorable mention heard on X.
11:53Uh, this is Gary Tan building a permanent memory for his AI agents that he gave it away, uh, and he's been adding more and more features to it. I think this is the from this week where he added voice, but I oh, wait.
12:06Where is it? Where is it? Uh, here.
12:09Let me go right back here because I want people to see what it looks like on the inside, and I've done an interview with the founders that showed tab here in AlphaClaw. I can show you my brain here.
12:18You can see that we've got companies. We've got we've got people. For example, founders, Adam Guild here created the world's first AI CMO for restaurants.
12:30See, so what he's got is a really well organized set of information on the people, the companies, etcetera, that he works with and and the timeline of what they did. That's what this is about, making your agent smarter by giving it an organized database.
12:45Peter, thoughts?
12:47Um, it's something that is very useful if you do have a lot of kind of disparate knowledge that you need to tie together in some way. So when it's someone like Gary Tan so he's the president of Y Combinator, for example.
13:00So he's working with hundreds of different founders all the time, needs to kind of keep track of, you know, who's who and who's working on what and that type of thing. I can see WireBrain would be useful. For me, I tend to be working on really, like, tiny type projects where I kinda know what's going on, so it's not so useful for me.
13:14But this is an interesting project. You know, I think it is put together by AI, but it's had some really good decisions behind it. So it's using Postgres as the database, which, you know, a lot of people will like.
13:26And it's doing some interesting stuff, like, uh, it's using embeddings, um, and BM 25 ranking for looking stuff up. So the the there is some quality behind it, but I've not really poked into the middle of this brain. I've not done any brain surgery on it.
13:40Um, but, yeah, people seem to be quite excited about this. They are. If you if you're working with a lot of clients, if you're on if you're talking to a lot of people and your agent needs to keep them all in order, this is a good way to do it.
13:51Alright. Uh, next, this one you asked me to add on. This is, uh, Webright by Microsoft.
13:57What is it?
13:59Uh, it's another project from Microsoft, which I always like to see. Um, it kind of builds upon their work with a project called Playwright, which is basically a library that, um, developers have used in the past and still still now, um, to do testing on web apps most commonly.
14:17But you could also use it for scraping the web and stuff like that. It's just like a a remote control for a browser. The the problem they're trying to address here is that if you wanna put an agent in control of a browser, you've got a few different ways you can do it.
14:29You can give it the access to the full screen view like some of these computer use things use now. But the problem with that is that while you get to see exactly what a user sees, you are looking at raw pixels, and that's quite demanding in terms of the amount of tokens going over the wire and the speed isn't that great and so on.
14:46Whereas Webright is much lower level, so that a agent can just see, like, the different elements on the page, but a more technical lower level like you would if you were programming something to remote control a browser behind the scenes.
15:02So their goal here is just to make the whole thing more efficient and significantly faster and more token efficient in particular. It's very early days for this, but I'm excited to see where it goes just because, again, it's from Microsoft, and they've had a a lot of success in this area with Playwright already. And this is I actually can see the web browser as it's using it, like in this demo that I just showed?
15:22Yes. But the a but the agent behind the scenes is seeing more like the tree of kind of what's on the page. Right.
15:27So it's not just it's not saying, like, look at these pixels and work out what's going It's more like, look at the code and work out what's going on, and then remote control it from there.
15:35Yeah. To create this document that I've been clicking through to show everything, um, I had to go and find links, and what I did was I asked Claude to go find the links, and I could see it saying, okay, I'm taking a screenshot. Now I'm understanding, taking a screenshot.
15:46That's what you're talking about. It eliminates it. And then the final, heard on x.
15:50This is really popular on, uh, on x, but not popular enough to make it to the top 10 list. It's, uh, Lightparse. What is Lightparse?
15:58I like this one just because it's a common task that we all have. We all have PDFs full of all sorts of things, or we receive them.
16:06They're full of media kits and company data and all this type of stuff. Um, and we need to work with them. And sometimes they have very weird formats and layouts to them.
16:14Um, so, you know, it'd be lovely if we could have everything in markdown, but sadly, we can't. Um, so PDFs will often be in columns or they'll have different tables in and things like that. This is something that, you know, without using AI can break down some of that stuff.
16:29Um, and we've had you know, there's tools that can do OCR and stuff in the past. It's just nice to see something that brings together the OCR and, you know, being aware of layouts and how to process PDFs that we can run entirely locally without, you know, having to shove all of our PDFs up to, you know, Opus or OpenAI or whatever.
16:46And it's on on my desktop. It doesn't need to have outside connection, and it just works. Yeah.
16:54Okay. Excellent. So unlike the other one that we saw, which does need API, does need outside access, this is light.
17:00It's on my computer. Works. Okay.
17:02Let's go on to the next one. Compound engineering plug in from Every.
17:08I like Every. This is a company that both makes software and writes content about how to make software with AI and how to use AI. I I like it.
17:17What is this, though?
17:19Um, I mean, I like Every as well. They do some great work, and they have so many different weird random little projects and, uh, services. I think you said you said you used one for looking at your email and stuff like that.
17:30And the design that they have is amazing as well. This is largely a giant set of skills, which tell, you know, all kind of instruct Claude and Codex.
17:41And they the way they've coded it is actually kind of cool because it actually plugs into almost any of the main sort of, you know, harnesses. And it includes things like, you know, how to do good Git commits.
17:52It includes things like even how to program Ruby on Rails the DHH way. Mhmm. And the thing I like about these types of repo isn't necessarily just to install them and just roll with it, because I don't necessarily like other people's opinions to take over how my agents work.
18:06But I like to dig in and actually look at what the skills are and see, is there anything in these skills that I can use? And I provided you special access with the directory to show you what those skills are, and it covers all types of things. You know, from front end design skills to how to think about projects and stuff like that.
18:23I like to dig through these and see if there's any gold. This is where it is.
18:26It's under plugins, then compound engineering, and skills you showed me before we got started, and then in here, I can actually see the skill. And what you do is you look through and you say, okay, how are they thinking about this? How are they thinking about, uh, commit?
18:39How are they thinking about each one of these and debugging? And that's what you want out of it. Absolutely.
18:44Because I don't necessarily wanna trust someone with a bunch of rules, and then just have them put them into my system, and I don't know what they are. I need to look at them anyway. So I might as well learn from the skills rather than just blindly trust them.
18:55And that is actually quite relevant for one of our future items that we're gonna cover. So, you know, you can't just blindly trust skills just jammed into your context from my perspective.
19:04And I've got a video here that I'll include in the in the report that everyone can get in the description where you can see one of the founders within the every community going through and explaining how this works, how he uses it, how they think about compound engineering.
19:20Alright. I'll let you all watch this afterwards. Onto the next one.
19:23Stop, slop. This is again last week, it was on number. It was top 10.
19:27This week, it's top 10. And what it does is it says, look, we all know when AI writing is AI writing. There are certain tells.
19:35Why don't we just tell the AI to stop using these tells? I think a great example of it is right here actually, not this one. The live demo where they said, look, These are the things that we know.
19:47Like, starting off with, here's the thing. Just get rid of it. It turns out, just get rid of it, and there's a few different things that it will eliminate.
19:55Would you use this for writing?
19:58No. No. And this is the one no.
20:00This is this is the one that I was foreshadowing, actually, when I said about those don't trust skills straight out of the box. You need to write your own one of these. And the problem with stop, slop in particular is there's almost nothing there.
20:10Like, if you look through the the documents, it's absolutely tiny. It it's just a bun it's just like a bunch of small rules, and some of them are very bad.
20:19So for example, there's one rule in it that says, if any sentence starts with a a question that starts with w h, rewrite it. So you can't start a sentence with what or where. It says any adverbs kill them.
20:32Well, adverbs can be bad, but I wouldn't necessarily say just get rid of every single adverb. So I would look at this skill and say, yeah, wanna take this.
20:42I wanna take that. But I wouldn't just use it blindly because it could just completely neuter, you know, your voice and your writing anyway.
20:50I mean, I know this is the process AI writing, but, you know, do you wanna take everything out of it?
20:57Yeah. Uh, it's okay.
21:01I'm in a hotel, and, uh, I guess checkout is a 10 because they just came in to clean up in here. I didn't know. I'll talk to them later.
21:08I thought you were getting a nice steak delivered. Where did you where did you even see that, the WH part?
21:14Uh, if you go into skill.m I think it was no. Actually, skilledmd just replies just responds to the other skill files.
21:22Um, I don't know where it was. I just wrote down notes when I was reviewing it. Um, but it's in one of those markdown files.
21:28But so, yeah, it was just like, get rid of this and get rid of that. Um, there are other ways you can look into this, just to let you know. So anyone that's watching this, if you Google for Wikipedia signs of AI writing, there's another really good document that Wikipedia has put together, um, which also has tons of telltale signs, which you might also wanna turn into your own skill file.
21:48I see. So you're saying create your own skill file. Don't just use theirs, and you can you can use theirs as reference, but not right out of the box.
21:58Yes. And, of course, models keep changing. Mod you know, you might be using a model tomorrow.
22:02You might be using Opus 4.9 or GPT six or whatever it is, and it might have these different tells. Well Right. That old document isn't gonna work.
22:09So you need to keep on top of it for yourself.
22:12Okay. Alright. I like that warning.
22:15Next is super memory. Every new chat with Claude or ChatGPT is like meeting a brilliant employee with total amnesia. You're re explaining everything every time.
22:24This gives your AI a long term memory, follows you across every tool.
22:30What do you think of this? And I've got someone here, uh, Ben Sigman, who actually evaluated this compared it to others. What do you think of this before we go into his evaluation?
22:39I have not tried this, but it does take us back to the g brain item somewhat in terms of how it works. The big thing that Super Memory is doing, which is a bit different, is that they're running it as a commercial service.
22:50So it's a lot easier for people in companies or whatever just to plug in and use it, but you can also use it in the open source form for free too, which is nice. So I like the way they've structured it from that that perspective. Um, but, again, in terms of actually using it, it's not the sort thing I necessarily would use just because I have my memory structured already in a in a different way.
23:08But How do you structure your memory? How do you avoid the total amnesia and or just, like, one time it knows it, the next time it doesn't? I like to start my sessions with the amnesia because I like to point it to things that we've done before as we go.
23:24I don't like to just say, just here, I have access to everything. Just go at it. I like to be selective about what I give it because I don't like it when so even something like Claude or ChatGPT, I don't like it when I ask a question and it uses my prior memories as like, oh, why are you asking that question?
23:41Because previously you did x y and z. And it's like, no, I'm asking this independently of what I previously discussed with you. Um, so I don't want that in my coding either.
23:51Alright. Fair point. I've got a video here explaining how it works.
23:54I like your feedback on it. Let's go on to the next one. Um, ECC.
23:59This is a builder who spent ten plus months living in AI coding tools, packaged everything he learned into 63 specialized agents, 249 skills into one free install. It's the closest thing to an operating system for your AI coding assistant.
24:12It's got does it really have 205 stars? And it's one of the biggest reviews on GitHub. Does it?
24:18Does it?
24:19It yeah. It's even more than 10 Yeah.
24:23And 31,000 forks, and still you have a controversial point of view on this. What is that?
24:30Also, did you Actually, mean I'm sorry. First, explain explain it in your words. What is this?
24:34And then what's your what's your thought on it?
24:37It's a giant pile of stuff. Mhmm. Yeah.
24:40I'm not gonna use any offensive terminology. This is clearly you know, someone has spent a lot of time on this and put something together. Problem is it's just too much.
24:47It's like, you know, weeding in a Boeing seven four seven or something just to, you know, cross the street. Do you do you trust it? Are you gonna read for all of it?
24:56Now, clearly, there's enough people that are lazy enough that they're just gonna go, oh, great. Loads of stuff. Download, click, star, and everything.
25:04But the fact is, I have no idea what any of this is doing or what it's been directed to do. So why am I just gonna trust out of the gate? Now the thing about the the one that Every did earlier is that at least with Every, you can trust them.
25:16They they're you know who they are. They have a established background, and they are experts in using this stuff. But how am I gonna trust this to know how to review Go code or Java code or Python code without going through it?
25:27And then by the time I've gone through it, I might as well have made my own. So I don't see the appeal of these types of things. It's too much all at once.
25:35I don't either. I could see him wanting to share, and other people may be wanting to learn, but it's just a lot a lot of information in here, a lot of different tools. I've got a video of somebody here actually using it.
25:45I'm gonna won't even play the whole thing, but I'll give people just a taste of actually, I don't even think it's worth even the taste of it.
25:53You'll have the video so you can watch it on your own. Basically, he said was he said, I want to add a, uh, calendar link to my site where people can book meetings with me.
26:01He I don't know why he needed the skill in order to do it, but he used the everything Claude code plan. He planned it, and he got the Calendly integration in there. Again,
26:13there I think these things are great. I think these things are great for the people that made them, because they made them the way they want them to make them, and they use, like, their their own like, their your if I took my approach and put it into a set of skills, then it works great for me.
26:27Doesn't necessarily work well for someone else. So you really need to review these carefully before you go whole hog on it.
26:33There's a Calendly. Yeah. I agree.
26:35Okay. Next. Um, and I I do think also it's nice to look over somebody's shoulder and see how they do things, see how they structure things.
26:43Yeah. But like you said, just using it right out of the box. No.
26:47I like this whole style here. Here's the problem that TasteSkill addresses. We now see every time somebody gives us a document that they think looks beautiful, we see that it looks just like Claude because they're using the Anthropic design.
27:00It drives me freaking crazy. Honestly, I'll be honest with you, Peter. These, like, web pages that you're all looking at right now, these all looked exactly like Claude.
27:08What I did was I couldn't stand looking at them after the first time. I then made some adjustments myself, and then I saw a repo in one of our past sessions which said, you can take designs from all these other companies and use them. So I did.
27:20This is from Intercom, the the software company. I use their design, and then I change the colors to make it feel like mine, and it worked, and it made it feel a little bit better for me. I'll probably keep tweaking it.
27:31And what I like about this is that it's doing the same thing. It's saying, use this to change your designs. We've got a few we've got a here, this is a video of a creator who tried it, and he's gonna show what the designs look like, and it's, of course, on the GitHub repo also.
27:47What do you here it is. Look, this definitely doesn't look like Anthropix design taste. No.
27:51What do you think of this?
27:53This is an area where I'm a little bit more inclined to go along with, you know, having a bunch of different skill files brought into a project. And the reason for that is that because it's about taste rather than how to do things. Mhmm.
28:07And so I don't mind other people's opinions necessarily being brought into design if you like the output. Because the thing is, like, you don't like what it produces.
28:16You can say, oh, screw that, and then let's go with a different approach. Right. And so this this project does build upon previous kind of steps forward in this area, like impeccable dot style, which I saw actually in the screenshot for that video, um, is another thing.
28:30But the thing is once you design with these skill files, often you will then start to see the same, like, tropes coming out over and over and over.
28:39So again, this is a situation where I think it's good to see what other designers' taste is Mhmm. But then build on it for yourself. Build your own things like this.
28:47And and Right. You know, the models now are actually so smart. You can even do cool things like if you see a photo that you like.
28:53So for example, I took a photo the other day of the Backrooms, you know, that's big at the cinema at the moment. And I said, design a web page that looks like the Backrooms. And threw it into Opus, and it did it.
29:02And it's like, oh, actually, this is kinda cool. Like, it's got the color scheme and, you know, all this type of you can do that now. So you can just find an aesthetic that you like, throw it at a model and just do it.
29:11You don't necessarily have to say, design this whole thing from scratch for me. You can give it inspiration. But if you wanna mix in things that, you know, give you, like, skills like this, give ideas on typography and fonts and stuff like that, blend it in, give it a try, see how it goes.
29:26Um, you know, I think these things are kinda cool.
29:29I do too. I like them for just seeing a different take on what I'm already doing. Alright.
29:33Yeah. Next. Understand anything.
29:36Here's the deal. Maybe you created something a few months ago and you forgot what you got there, or maybe you started a new project and you're looking at something for the very first time. You've got hundreds of thousands of lines of code.
29:46You don't know what anything what anything does, what it means. What this will help you is understand it by turning it into a visual that lets you see what's what. That's what understand anything is.
29:57What do you think of this?
29:59I've not used this one, but it does plug into an idea that I've had for a while now, which is that try and avoid using LLMs for writing too much and focus on using LLMs for reading things because they are much better at that job.
30:14So I would be much more likely to take a bunch of skills or, you know, plug ins like this and have it process something that already exists because it can extract insights from things perhaps better than it could recreate that thing.
30:27So, yeah, I have absolutely no problem with projects like this. And there is a lot to reading large code projects that you can't just go to, you know, Codex or Claude and say, I'll read this project and tell me x, y, and z. Sometimes there is a lot of analysis to be done, um, on top of it.
30:43And if these skills give it the extra power to get over those hurdles, then it could make the analysis even better. So yeah.
30:51But again, I've not used this one, so I don't know how well it does. But a lot of people seem to like it.
30:57There's a little bit of controversy here on hacker news. What what could people be upset about with this?
31:03Uh, was there a thing about, uh, yes. Someone was saying that the stars might not be, um, entirely legit, which is a common problem on GitHub. You know, there's this whole thing about that you can pay, uh, to get your repo starred.
31:17And, you know, the reason for that is that it's the modern version of kudos. You know, it's if your repo has 10,000 stars on it, then it must be better than one with a 100.
31:27It's not necessarily true, but it is a you know, it's a hurdle to overcome. Um, I don't know anything about this project and whether it's done that, but that is something that I think people in that hacking news thread were thinking might happen.
31:38think we need a better solution for this. We need something like, what do people who I care about think of this? Or how many of the people who are creating give it stock?
31:47I don't know. Something. Digg is is messing with this a little bit, digg.com.
31:52I don't think anyone's nailed it yet, and it's bothering me and it's hurting the whole ecosystem. This one you love. You actually told me, Andrew, Yeah.
31:59Put this VoxCPM. What is it?
32:05Okay. So we've had a load of text to speech things over the years all the way from the eighties where we had, like, you know, robot kind of voices, um, through to the sort of systems that Stephen Hawkins used.
32:15Um, but then in the last five to ten years, you know, AI has really turned that dial up from, like, a two out of 10 quality to sort of eight out of 10. Um, and this is actually pretty good.
32:26Um, you know, at first, it was I was saying 11 labs quality, like, that's what they're claiming. Right. Um, I would say it's not quite there, but it's, like, 95% of the weight of that.
32:36Um, and it was so good, actually. I could run it on my own Mac. Like, it downloaded, compiled, downloaded the model, and I was up and producing audio with this, you know, within ten minutes.
32:46And the output's not bad. And the good thing is you you can sculpt it as well. So you can say, I want a voice that sounds like
32:52x, y, or z, and it will do that. I'll do it right here. Uh, old man with raspy voice, and then I could have him say, get off my lawn.
33:04And this is not working right now on, uh, on Hugging Face. Uh, but you get a sense of how it is. You can also upload your own, uh, voice as a reference and allow it to take off from your voice.
33:17I've got a video here to show how it works. Um, it's it's been amazing.
33:22I don't know how Eleven Labs is gonna compete when all this stuff is so is so free. It's becoming so accessible. People are are able to use it, like you said, on their own computers.
33:31Super easy. What do you think what? I don't know where Eleven Labs is gonna go with this.
33:36So well, the one thing that Eleven Labs has going for it is that they've really focused on having professional voice actors actually supply their voices with permission, with licensing. So that's the win there, is that companies that wanna have voices that every single legal thing is ticked and crossed can go to Eleven Labs, they provide that kind of compliance side of it.
33:58Um, it's not necessarily just about producing speech. It's about producing a paper trail, which, you know, actually, I'd realized.
34:05I just sounded like Claude then the way that I explained that. Um, but that is basically what it is. Yeah.
34:10Exactly. So the thing is, you know, if you produce audio with this, it could be cloned from anyone's voice and, you know Right. How do you know what your system's doing in the background?
34:19You know, it may have downloaded a video of, you know, some famous politician and just cloned a voice for you without knowing it. So,
34:24yeah, Eleven Labs isn't gonna let you do that. Here. Let's hit play on this video, which, of course, we'll have in the report for everyone who wants it.
34:32VoxCPM is an innovative end to e beyond the ancient oaks and the misty trail. A secret remains on I'm convinced that Dan from Smart Tutorials on this video is actually just a voice using VoxCPM.
34:46Listen to this. Like, look at this. I actually think that he's this is not the case.
34:50Today, we are diving into VoxCP I don't know. I don't know.
34:53I'll let people decide for themselves. Alright. That is number 10.
34:58We've got the full report in the description, you can play with this, listen to it, try it yourself. And, Peter, where do people see you?
35:05You can find me on Twitter. Oh, that's not Twitter now anymore. It's x, isn't it?
35:08Cooper x eight six
35:10or p two c dot org on the web. And, of course, we'll have a link below. And now that you've listened to this, we've got 10 other repos that were super hot last week plus three special ones that you've gotta see.
35:19There's a link right here on the screen. Click it, and we'll see you over there.
The Hook

The bait, then the rug-pull.

The title promises 95% LLM savings and an ElevenLabs killer. The show mostly delivers -- just not in the way the headline implies.

Frameworks

Named ideas worth stealing.

20:00concept

Skill file trust hierarchy

Before installing any AI skill bundle: (1) Is the source a known trusted team? (2) Have you read every rule? (3) Do the rules match your opinions? Install only what passes all three.

Steal forAny post about Claude Code CLAUDE.md or AI agent configuration
03:13concept

The real cost of a proxy layer

A compression proxy reduces tokens only when traffic is genuinely bloated. In ordinary sessions the proxy overhead can exceed the savings.

Steal forAgent architecture decisions
CTA Breakdown

How they asked for the click.

VERBAL ASK
34:57link
We have got the full report in the description, you can play with this, listen to it, try it yourself.

Soft close -- full report URL in description, no hard sell. Peter shares his X handle.

FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
Storyboard

Visual structure at a glance.

hook teaser
hookhook teaser00:00
MoneyPrinterTurbo
valueMoneyPrinterTurbo00:36
headroom claim
valueheadroom claim02:42
debunk 4.8%
valuedebunk 4.8%03:21
markitdown
valuemarkitdown04:39
Zapier sponsor
ctaZapier sponsor07:03
Odysseus PewDiePie
valueOdysseus PewDiePie08:18
compound-engineering
valuecompound-engineering17:06
stop-slop
valuestop-slop19:21
ECC controversy
valueECC controversy25:57
VoxCPM
valueVoxCPM32:06
wrap-up
ctawrap-up34:57
Frame Gallery

Visual moments.

Watch next

More from this channel + related breakdowns.

Chat about this