Modern Creator
Jack Roberts · YouTube

9 Insane GitHub Repos to 10X Claude Code

Eleven tools, each one tested against plain Claude Code, with the catch for every one spelled out: free models run slower, skill packs cost more, and design kits trade one kind of slop for another.

Posted
2 days ago
Duration
Format
Listicle
educational
Views
22.2K
355 likes
Big Idea

The argument in one line.

Most Claude Code add-ons trade money, speed, or quality against each other, so the right repo depends on the job: free models for small tasks, lean skill sets over giant packs, and cheap context for web reading.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…
  • You already pay for Claude Code and want to stretch your usage limits further without dropping to worse results.
  • You build sites, copy, or videos for clients and want output that does not read or look like default Claude.
  • You are deciding whether to install a giant skill pack like Everything Claude Code and want real token and cost numbers first.
  • You want an agent to control desktop apps or read websites without burning tokens on screenshots and raw HTML.
  • You are curious about running open models inside the Claude Code harness, including free or uncensored ones.
SKIP IF…
  • You want deep setup walkthroughs; each tool gets a few minutes and the install steps live in a separate downloadable pack.
  • You are brand new to Claude Code and have not installed it yet; the video assumes you already use the harness daily.
  • You need production-grade reliability guarantees; most tests here are small, single-person benchmarks.
TL;DR

The full version, fast.

Every Claude Code add-on is a trade, and the test numbers show where each one pays. Free LLM API runs the Claude Code harness on 87 free providers for $0: it passed all 9 small jobs but ran 2.6 to 4.6 times slower and passed only 2 of 3 big ones. Agent Desktop drives apps through the accessibility tree instead of screenshots, and Firecrawl cut one page from 58,800 tokens to 2,900. Big skill packs like Everything Claude Code made chats 63% heavier for the same score, so load the seven useful files instead. Use structure-heavy skills on cheaper models, run copy through an anti-slop pass, and treat design kits as a start, not a finish.

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.

00:00 – 00:30

01 · Cold open

Nine repos promised, from free Claude Code to uncensored models.

00:30 – 05:29

02 · Free LLM API: Claude Code for $0

Run the Claude Code harness on 87 free providers. 22 runs across 4 jobs: all small jobs pass, 2.6 to 4.6 times slower, 2 of 3 big jobs pass.

05:29 – 07:35

03 · Agent Desktop: buttons, not screenshots

Computer control through the accessibility tree, faster than screenshot loops and usable by text-only models.

07:35 – 11:08

04 · Firecrawl: cheap web reading

Brand extraction, competitor analysis, traffic data, and asking many sites one question, at 2,900 tokens instead of 58,800.

11:08 – 14:30

05 · Lemo Opuscar: film with no camera

99 motion-graphic film styles with finished samples, plus a community pitch and the token cost of its built-in revisions.

14:30 – 18:10

06 · Cardboard: AI video editor

A YC-backed editor where the whole edit lives in one text file, edited in plain English, with switchable models.

18:10 – 20:49

07 · Agent Skills: Claude as a senior engineer

25 skills across spec, plan, build, test, review, ship. Same score on small jobs, four times slower.

20:49 – 22:54

08 · Everything Claude Code: 314 skills vs 7

All configs find 20 of 20 security holes; the full pack makes chats 63% heavier. Keep the inspector and the diary.

22:54 – 25:18

09 · Slop Monster: copy that reads human

Five clarity checks plus marketing rules strip AI tells from landing pages, cold emails, and support replies.

25:18 – 27:18

10 · Astryx by Meta: anti-slop design

174 components and 58 templates remove the default Claude look but overwhelm; Power Design as a quick fix.

27:18 – 29:13

11 · Claude Code Game Studios

Brief, build, check: a far better game than one prompt at about 20 times the cost. Prototype or 49-agent studio mode.

29:13 – 30:50

12 · Heretic: uncensor an open model

Find and remove the refusal direction. A small Qwen model drops from 92 refusals in 100 to 4.

Atomic Insights

Lines worth screenshotting.

  • A free model inside the Claude Code harness passed 9 of 9 small coding jobs but ran 2.6 to 4.6 times slower than Claude.
  • On bigger coding jobs, free routed models passed 2 of 3 runs, and one timed out, so they are a spare brain rather than a replacement.
  • Performance splits into the model, which is the driver, and the harness, which is the car, and you can swap the driver while keeping the car.
  • Screenshot-based computer control is slow and token-heavy; reading an app's accessibility list of buttons skips the picture entirely.
  • Accessibility-tree control lets text-only models like DeepSeek V4.1 Flash operate desktop apps they could never see.
  • Fetching a page with curl cost 58,800 tokens; the same page through Firecrawl cost 2,900, about 95% less for the model to read.
  • Your AI only knows what it can fetch, so cheap clean web reading directly stretches your usage limits.
  • Video edited in a single text file is easier for an agent than clicking through Premiere Pro, where every button was designed for a human hand.
  • Senior-engineer skill packs scored the same as plain Claude on a small clear job but took four times longer.
  • Process-heavy skills pay off most on cheaper models; top models increasingly do the planning on their own.
  • Loading all 314 skills of Everything Claude Code made every chat 63% heavier and doubled review cost, with no gain in bugs found.
  • Plain Claude, the full 314-skill pack, and a 7-file subset all found 20 of 20 planted security holes; the 7-file version cost $0.22 versus $0.38.
  • AI copy is detectable because it is predictable: unproven numbers, 'not just X, it's Y,' and words like seamless and empower.
  • An anti-slop rewrite pass cut a landing page's word count by about 40% while making the offer clearer.
  • Six AI-designed landing pages made the same three moves: a colored phrase in the headline, a 'see how it works' second button, and a 'no card' line.
  • A design-system kit can remove the default Claude look while adding a different problem: pages so dense nobody knows where to look first.
  • A brief-then-build-then-check game studio produced a far better game than one prompt, but cost about 20 times more.
  • Abliteration found the refusal direction in a small Qwen model and dropped its refusals from 92 of 100 requests to 4.
Takeaway

Every Claude Code upgrade costs something. Measure it.

WHAT TO LEARN

Free models, skill packs, and design kits all trade speed, tokens, or quality, so match each tool to the job and keep only what earns its cost.

02Free LLM API: Claude Code for $0
  • Free routed models inside the Claude Code harness handle small coding jobs at $0, but expect them to run 2.6 to 4.6 times slower than Claude.
  • Keep free models off big jobs due today, walk-away tasks, and private client work; one in three big runs failed and one timed out.
  • Swapping the model while keeping the harness preserves your skills and tools, so a cheap model becomes a usable spare brain.
03Agent Desktop: buttons, not screenshots
  • Computer control through an app's accessibility list of buttons beats screenshot loops on speed and tokens, and works with text-only models.
  • When a hosted agent's own tool is blocked from an app, an accessibility-tree controller is a practical way around it.
04Firecrawl: cheap web reading
  • Read websites through a cleaner like Firecrawl instead of raw fetches; one page dropped from 58,800 tokens to 2,900.
  • Use web extraction for competitor research: brand palettes, fonts, pricing limits, and traffic, with a source page cited for each fact.
05Lemo Opuscar: film with no camera
  • A large library of pre-built visual styles fights the generic Claude look because you start from a distinct design instead of the default.
  • Skills with built-in revision rounds give better one-shot results but burn more tokens, so commit to a style before running them.
06Cardboard: AI video editor
  • Agents edit more reliably when the whole edit lives in one text file than when they click through software built for human hands.
  • Keep plain-English editing reversible: every change in a text-based edit can be undone instantly, which makes experimenting cheap.
07Agent Skills: Claude as a senior engineer
  • Follow spec, plan, build, test, review, ship for any real build, and define what done looks like for each small task.
  • Skip heavy process skills on small clear jobs; they matched plain Claude's score while taking four times as long.
  • Structured skill packs return the most on cheaper models, since top models increasingly plan well on their own.
08Everything Claude Code: 314 skills vs 7
  • Loading all 314 skills made every chat 63% heavier and nearly doubled review cost without finding a single extra bug.
  • From a giant pack, keep a checklist-driven inspector and a diary that records progress for the next chat, then drop the rest.
09Slop Monster: copy that reads human
  • Run AI copy through a check for unproven numbers, 'not just X, it's Y' constructions, and inflated words before it ships.
  • A good anti-slop pass also tightens the offer: it cut one landing page by roughly 40% and made the benefit concrete.
10Astryx by Meta: anti-slop design
  • AI landing pages repeat the same three moves, so changing those alone makes a page look less generated.
  • Component kits remove the default AI look but can overwhelm the page; follow up with prompts on typography, spacing, and hierarchy.
11Claude Code Game Studios
  • A brief-build-check loop produced a much better game than one prompt, but at roughly 20 times the cost.
  • Use a quick prototype mode to test an idea, and save the full multi-agent studio for projects worth the spend.
12Heretic: uncensor an open model
  • Hosted models can refuse harmless work like security hardening, so keep an open model you control as a fallback.
  • Abliteration removes the refusal direction from an open model; one small model went from 92 refusals in 100 to 4.
Glossary

Terms worth knowing.

Harness
The software wrapped around a language model that gives it tools, files, skills, and a working loop. Claude Code is a harness; the model inside it can be swapped.
Model router
A service that sends each request to one of many models, picking by availability, speed, or quality score, so you can use free providers through one key.
Accessibility tree
The structured list of buttons, fields, and labels an operating system exposes for screen readers. Agents can read it to control apps without taking screenshots.
Skill
A packaged instruction file that Claude Code loads to follow a specific workflow, such as planning, testing, or reviewing code.
AI slop
Generic, predictable AI output that people recognize instantly, such as stock phrases in copy or the same layout moves in every generated web page.
Design system
A shared library of components, page templates, and themes that keeps an interface consistent across many screens.
Abliteration
A technique that finds the internal direction a model activates when refusing a request and removes it, producing an uncensored version of an open model.
Open-weight model
A model whose trained weights are published so anyone can download, run, and modify it locally, unlike hosted closed models.
Resources

Things they pointed at.

Quotables

Lines you could clip.

01:05
“There's the model, which is like the race car driver, and then there's the harness, which is like the car it drives.”
clean metaphor, no setup needed→ TikTok hook↗ Tweet quote
07:38
“Your AI only knows what it can fetch.”
seven-word thesis→ newsletter pull-quote↗ Tweet quote
09:50
“If I would have just had to scrape the page, that would have been 58,000 tokens. With Firecrawl, that was 2,900 tokens.”
hard number contrast→ IG reel cold open↗ Tweet quote
19:31
“It's almost like sending a bazooka to open up a door.”
vivid image for overkill tooling→ TikTok hook↗ Tweet quote
21:29
“If you're loading in 314 skills, it's like double the price.”
contrarian take on skill packs→ IG reel cold open↗ Tweet quote
30:31
“It refused 92 out of 100 of my controversial requests. After I obliterated it, it only refused four out of 100.”
before/after number payoff→ TikTok hook↗ 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.

metaphor
These are nine free repos that make Claude so good it feels illegal. Which includes a repo that lets you run on the Claude code harness 100 % free. A huge upgrade to computer control.
A repo that's going to turn Claude into a senior engineer. Video editing upgrades, destroying AI slop, and even a model that lets you uncensor AI. And if you're new, I'm Jack.
I built and sold my electric startup with a ton of customers. And on this channel, I share the stuff that works. So if you haven't already, grab that beautiful coffee.
and let's dive straight in. So we have nine new repos. I have tested every single one of them, beginning now with level one, which is free LLM API.
Now, the idea here is that to code and build things, we need a subscription to Cloud or to Codex, right? Well, what this lets us do is use the Cloud Code harness. for free with basically free models now it gives you 87 different free providers essentially we just use this one repo here free llm api then essentially what we can do is use the skills of of the cloud code harness to create anything we want to because as we know when it comes to performance there's two things there's the model which is like the race car driver and then there's the harness which is like the car it drives so we get the exact same harness the cloud harness but we're using a free model instead and so for example this is a cloud code that I built with an uncensored model yesterday in the video.
I'll put a link on screen for you. And essentially what I can do now is come down here and pick if I want to free models, which is amazing. And now I can ask any question I want to like, hey, who was the tallest man to ever live?
Whatever it looks like. And then effectively, I'm now using a free model inside the Cloud Code Harness. Plus, I have access to all of the skills I have, just like in regular call.
Claude. And guys, this came back and this was 100 % free. As you can see, the cost is $0.
And this is not even running on my own computer. Though if I wanted to, I could actually go ahead and run one of those local models just as easily. And the way that it effectively works is you get multiple free brands.
And so for example, I just built a local environment with the Claude code harness that looks just like Claude. And now I can use any model I want to with it. And as you can see, it even built me a Python script with these completely free models.
But I actually went ahead and tested everything I'm going to show you in this video. So you can cut all the nonsense and get straight to the point i did 22 separate runs and did the same four jobs in both routes so one was like a free route with this free llm api another one was just with regular code and would you believe it i had to do a to -do script with tests tidy up multiple files and fix three bugs and do six tests now as you can see that both of them actually completed it but the difference is this column here cost me zero dollars but the difference is it probably took somewhere between like i don't know two to four times as much time then i had to do a small package for me and as you can see the Gold code was faster, but obviously was expensive.
But as you can see, both completed it. It just took more time with the free models. Then when it came to something slightly more complex, it actually passed two out of three of the tests, was called crushed all of them.
And the time -wise, again, it took longer to do. And essentially what happened is with a more complex task, it actually timed out. And in the second link of the description, I'm putting this nine free repo pack for you so you can follow along with this.
All you're going to do to come down to the relevant... section like cloud code for zero dollars i've included tips tricks full guides full screenshots for you all you're gonna do is come down click copy and then head over to cloud code and simply drop that in and just follow the instructions and this will walk you through exactly how to get the setup super easily and by the way these repos just get more and more powerful so you need to make sure you stick to level nine if you want to unlock the full capabilities with this now why did one fail well the way this works is it roots into a different model and some of them are quite strong and some of them aren't as strong as others which means that sometimes with big attacks asks, you may get some failures.
And using this, you can get capped with requests, for example, about 150 requests for the one key. And the way that this works, guys, is essentially you're going to pick a free brain. Again, if you want to access my systems, my prompts, I'll put a link down below for the community if you like.
But eventually, you're picking one here, and then the router itself is picking one of the different models it has. So you can see it's got dots, free note preview, Nemetron, 3 .5 lightning, north minicode, Nemetron, etc. It is picking one of them, and it kind of ranks them on reliability, speed, and intelligence, which ends up in they score.
And then as you get the app, you can kind of take the ones that you want. And then with me, for example, we did four jobs, 22 runs. As you can see, it passed all nine of the small jobs.
It was 2 .6 to 4 .6 times slower than Claude and only passed two out of three of the big jobs. And as you can see, if we run to 10x speed, you can see exactly just how long that actually took to go ahead and complete it. Now in terms of when they're good, they are good for this.
They are good for small jobs, okay? Because it crushes the honor. It's good if you just want to try out ClaudeCode.
You do need the $20 ClaudeCode subscription just to get started. It's good for chats and ideas it answers super quickly it's kind of like the similar performance as you'd get from like a local model it's good for learning and demos and a spare brain when you want it but then firstly what is it not good for speed okay it is 2 .6 to 4 .6 times slower i didn't have a mustache when i started using some of these models big jobs that you need today probably not the most reliable thing for it it's good if you know you're trying to do some coding and you want to continue and you're not credit you can just use some of these three models and again it's not so good though however if you want to leave it alone and come back to a completed task because of a trust element to it and not something that you'd give to your private clients now to use it head over to this guide like so click on repo one come down and all you're going to do guys is come to this first thing here and copy this once you've done that guys you're going to end up with a page like this once you sign in and create an account and like i say you're going to come down and you're just going to tick all the things and you can see
everything from your token budget and everything else that you need. Then we need to point Claude code to it, which is the next step in the guide. And once you've run this and pasted this into Claude, you need to come down right here.
paste that into Claude, and then you can actually come down and use these prompts or whatever you'd like to. But remember, always, always, always test these models. Now, free call code is one thing, but it's nothing, guys, when you think about the fact that we're still kind of stuck in the terminal with some of this stuff.
So this is where agent desktop is going to come to the rescue here. Now, this essentially is a completely game -changing thing that's happened with AI, and not many people have really started using it now. And this is the idea that our agent, typically speaking, can be stuck in a terminal.
But where it becomes really powerful is its ability, guys. to actually do computer control. Now, the way that most computer control works is it looks at a picture, it translates that into text, and then has to approximate back and forth.
Now, this is token intensive. It can sometimes take control of your mouse, but actually what we can do with this. is look at it from a list of buttons using the sort of accessibility controls.
But with this repo, guys, we can actually use it as a list of buttons instead. So it doesn't need to take a screenshot. And so, for example, this is a test I ran with Claude by itself, Claude with Agent Desktop, and then Codex with Agent Desktop to see how long it would take it to do a calculation in each of these things.
And look at this, 7 .9 seconds, Claude with the repo absolutely crushed it. So why should we use this? Well, sometimes if you're using Codex or Claude, it will actually get blocked sometimes.
This enables you to do it and oftentimes can do things a little faster. Now to do, you're going to click the link down below to install Agent Desktop. It's super duper easy to do.
And then you just ask it to do any job you want. But where this becomes really powerful, guys, is the fact that you can use Agent Desktop with any model. So say, for example, that you want to use...
deep sea v4 .1 flash our favorite model right we can now give it this repo and get it to do screen control even if maybe not necessarily had that because whilst it's not using the vision technologies and controlling your mouse we can now use this as a route to do other kinds of screen control thing which effectively give that giant beautiful chinese model hands to do things on your computer which is incredible and we can do it inside the cloud code harness too.
But one caveat with this, guys, is that sometimes cloud itself can be a little faster. So generally speaking, it really depends if it's a technique that you're going to be doing more and more often, which leads on to the problem, right? Because your agent can click any app, but when it comes to clicking on a website and reading things, not only is it slow, it's also very expensive, which leads us very nicely onto level three, which is Firecrawl.
Now, the issue that we've got, and not many will realize this, is your AI only knows what it can fetch. What do I mean by that? We want to use something like Firecrawl.
Now, Firecrawl itself has a repo. It's pretty exceptionally powerful. Now, the way this works is you add Firecrawl to anything like Claude or ChatGPT.
It is one connection, unbelievably simple. Because if you're not using Firecrawl or something like Firecrawl, you are going to be wasting money and you're not going to be getting the best level of intelligence that you can be. Let me show you exactly what it did.
I asked Claude to go ahead and use the Firecrawl API to go ahead and do some intelligence. Now, not only was it able to... to go ahead and get screenshots of everything.
So Gladys, my speech -to -text startup, for example, it read all the pages. It did a full brand extraction. Look at this.
The level of detail I got with this, it got the color palette, it got the background, it got the text, it got the logo, and it did it at a fraction of the price. And it was incredibly fast. I even got individual assets from, as you can see, I've got this beautiful interactive map.
I just said, hey, use Firecrawl. This was built with Firecrawl. Go ahead and do it.
Go and scrape everything from my brand and go and find competitors. And you can do the exact same thing for your business or your clients. And this even came back with a voice analysis, your world's fastest tool, dictation tool.
I'll also explain why this is so important to understand shortly as well. We then checked it against a live page. It found different things here in terms of the fonts and the nicknames and the different aspects like that.
Then it went ahead and did a full competitive analysis. And what was really cool about this, guys, as well, is we went ahead and got the typography of different competitors, found out, for example, their turn. And let's say I wanted to do a brand analysis or build up something interesting.
You can do it with that. Again, it did one for Aquavoid. I get a view, really nice crisp thing.
And as you can see, I've just got all this data that I can chat to about my team, which is awesome. We can see just how much slower they are than the wonderful Glido. Even guys, which is really cool, got traffic breakdowns, like how much traffic are they getting, which is really cool.
And we can see where the visits are. And again, it's just come down and explained and compared all this data for us. And you just imagine giving this to one of your clients or just having this like crisp, beautiful analysis.
But the cool thing about Firecrawl is you can actually... Ask websites questions. So for example, I said, which of these five dictation apps has a free plan?
And what is the exact limit of each one? Give the page for each fact. Do not guess.
And it came down and told us exactly what the answer is. And why is this so important, guys? If I would have just had to scrape the page, that would have been 58 ,000 tokens.
With Firecrawl, that was 2 ,900 tokens. So you want to think about how do I get more use of Fable 5 .2? How do I make my usage go further?
Using an app like Firecrawl is exactly how you do that. And we literally, guys, saved nine. 25 % on all of that information.
And just think about how much time you do that, which is why this has to be on the list. Some of the things that we can do, we can use it to size of a competitor, how many visits do they get? We can copy brands.
We can repricing. We can ask many sites one question. So we can just ask a question, which is cool.
We want to check out new markets. We can do that. And we can even build lead lists with it.
So the first thing to do is click the link down below. I'll put all these links for you in the pinned comment. I'll put five crawl down there.
So if you click that link, it will take you through to this page. And the first thing you've got to do is create an account then once you've done that you're going to head over to claude code come down click this button here and then click on connectors and then you're going to scroll up here guys to browse connectors and once you've done that you come to the search bar and you're going to type in fire crawl and just like that it'll pop up you're going to go ahead and click on connect it will open up a brand new page click on continue connecting and then you're going to see this page like so come down click on authorize the roi on this is spectacular they also have a paid plan i currently use it i find it incredibly helpful so claude now has access to the research But that's not the same, guys, as having a video, which is why Level 3 is a Rupert Cool theme.
A product video is something that we can use for a million different things. Now, the usual way, of course, is to use motion software and set that by ourselves. And by the way, if this all sounds like I'm speaking Spanish, I'm going to put a link down below for my community that has access to this agentic operating system that will get you light years ahead of Cloud Code.
Inside, you can learn everything from foundations, this Cloud Code course, which I update all the time, the agentic operating system, which has access to everything. all these beautiful systems that go from cloud to local. You're going to learn incredible skills like Crockbar, Content Creation Studio, and it will get you light years ahead of everybody else.
If you want to come check it out, I'll put a link down below so you can absolutely crush it. Now, here's the thing that you need to kind of think about here. This is a video that was designed with this repo.
Now, we all know, for example, that Claude is really good at motion graphics, but the question is always, well, what design should I use? How should I do that? The really cool thing about this repo, and again, this is just built...
off a single prompt trying to explode what glider does now it's quite a funky quirky style it looks pretty decent this repo is very very cool and just gives you a lot of different styles and designs that you can use. And then what you get with this guys is like 99 separate different styles.
As you can see, this is the one that I use when I pick Lido. You can basically just pick anyone that you want to. And the question is, why don't you just ask Claude for a video?
Well, a couple of reasons really. One is that the look is kind of solved. Like the more that people use Claude, the more it will have this Claude -y feel, which is why you'll often see that people impulsively sometimes say stuff like, that's AI slopping.
So what do you mean by AI slop? They can't tell you, but what they're actually saying is, this is something that I've seen everywhere and i can't quite explain why i don't like it and this is because people aren't using unique design so you need to use a repo like this and one that i'm going to show you soon it's going to get you make sure you buckle in that's what i'm going to tell you so look you can see it first every cell has a finished simple sample which is basically going to reduce the number of times that you need to go back and forth with these to find a style that you actually like which i think is pretty cool to be fair you know and you may not like all of these styles but some of them at some point are probably going to be cool right like I don't have a horse and race for this.
If it isn't good, it isn't good. But my opinion on this is like, it's good to have a full library on here. Like for example, inside my agentic operating system, if I show you in design, I built one of these out because I thought it was really important.
In my design section of motion here, I actually myself went ahead and built loads of these different designs that are to my particular design inclination, you know, to help my community out. Because I think like whatever you're doing, you do want to build out a series of beautiful motion graphics. And this one here that I'm showing you is an open source.
one that you can grab for free so you need to do guys is bring any of your own footage and logo logos or voice or music you don't get any of that stuff and again guys to download it click on repo number three come all the way down here and as you can see you can come down see all the benefits what it's for how to set it up you're just going to copy the skill you're going to show up to claude you're going to paste it in and then you can actually go ahead and do it and again i've kind of outlined when to use it and when not to use it for example one catch with this one guys i will tell you right now is Don't use it unless you really like the style that you've got.
Because what's built into the skill is multiple different revisions, which basically means that the one shot you get is going to be better. But at the same time, it's going to cost you more tokens at the same time. So you just want to make sure that, dude, I really, really freaking want this mid -century cartoon.
Just make sure that's the case. But it does also mean you can use it with like an Opus 5 .5 instead of like a Fable 5 .2. Now, that is filmed with no camera.
But the footage that you did film still needs cutting. And a lot of us have many subscriptions. editing software, which is where the next one actually comes in here called Cardboard.
Now Cardboard itself isn't actually a repo, but I was so impressed with this. I had to include it. It's a Y Combinator backed business.
And let me explain to you why this is so cool. The issue we have is editing takes a long amount of time. And whilst we can edit in these apps, what you really want, if you want to get ahead, is some specific software that is a freaking expert at the thing.
So for example, a lot of us are just having AI click through Premiere Pro, myself included, I used to do that, or whatever app that we're using. And the thing to bear in mind is that in Premiere Pro, all the buttons that you see are designed for humans. They're not actually designed for AI.
And when it's clicking, it's guessing to some extent. But with Cardboard, the idea is that the whole edit is in a text file and every change is unbelievably easy to undo. So if you made a mistake in Premiere Pro, you've got to come and move things around.
In Cardboard, it's... just like that. And it's back.
And I will tell you that the actual UI, when you sign up guys, it's pretty freaking cool. They put a lot of effort into the signup journey and I am super impressed with it. Now we download it.
It's a very, very super slick experience and you can get started by the way. For free, I'll put a link down below. You've got to come and check this one out.
Now, we're on the dashboard, and essentially what we can do is kind of just describe the thing that we want, and we can drop in any footage that we like here. And the way it works is we just talk to it in plain words. We can change our mind whenever we want to.
You can try any ideas, and it's all managed in one text file. So essentially, Cardboard, think of it as the place where the editing happens, and then you can link in with this either chat GPT or code. And what's even cooler, you can link it with something like Higgs field for image and video generation, link that to Cardboard, and then let you build anything.
So to use it, step one is just upload the footage just like you would do in... any editor and as you can see you get a really cool timeline down here in the bottom this is kind of similar like any other editing software that you'll actually see like premiere pro etc etc once you've done that we're just going to explain with english so i actually remember one christmas i locked myself away for days trying to figure out editing it would have been so much easier and cooler if you could have just done this to be fair so i can say hey make me hold the end card hold two seconds longer whenever i want to now this is one that i made completely with cardboard I had Hicksville generate the videos and then it generated the graphics.
I basically just described what I wanted. And oftentimes when you're editing for clients anyway, you are really just explaining the thing. So now I have this whole thing and Hicksville generated the videos at either side of it, which is really cool, but you use any whatever generation.
But the key hero here is the fact that it's all being edited in code in one place. And all I did basically was explain what I wanted to do. And so inside the app, for example, I can come down and click on this, which is cool.
And I just have this nice timeline. And look, we've even got some music on there. guys we're everything we want just out all right i like it how about this hey dude i like it one thing i want to change that instead of having start free with the green button can we just change that to a red button uh that'd be great thank you okay beautiful can you see just the interface is just really crisp and nice it's just really freaking crisp and nice you can drop videos in here do whatever you want to and i will say this that bam it's done it's done like It just trebled back in time to make that happen.
That was pretty freaking quick. I like that. I can come down here, pick my models, what I want to talk to.
One thing I'll tell you I really liked about this is when I actually signed up, it automatically identified what models I had on my computer. It knew that I had ChatGPT. It knew that I caught.
And the cool thing, honestly, is the fact that if I'm running low on failable credits or Opus 5 .5, I can just switch over. Yeah, let's bring over to GPT 6 .1 Terra. You know, that's a Terra job.
And so I can have Terra running and doing all the nominal stuff and then bring in the heavy firepower. power when i want to take it to a new level honestly a really tidy little lap and i wanted to include this because we talk a lot about motion design and stuff like that but honestly just a nice place to get video stuff done is really cool and then we also have these layout presets we can drop in um various different reasons you can see and it's like it just gives you the ability if you want to jump in and move things around like let's say i don't want this founder in the morning anymore She's gone.
She doesn't have to be here. It's all complete. And so now we've got editing locked down.
The next thing we need to do is make Claude act like a senior engineer. And the way we're going to do that, guys, is with a very specific set of skills. Now, these are 25 skills that make it plan before it goes ahead and does it.
So I've got here a... 25 set of skills, what each one does, and a prompt that you can copy right now to get started with this. And it's a group called Addy Osamny.
It's very cool. It's got six stages, and basically you can pick one to see what it builds. So I've got, if I want to define, I want to build, I want to test, I want to review, I want to ship, I can just go through and pick anything that I want to, or maybe I want to look at all of them together.
Let's say, for example, that I want to go ahead and plan something. Well, look at this. We've got slash plan, which cuts the job up, turns the spec into small tasks in order, and each task has its own done test.
And remember, one of the best ways to get the most out of the model is be really specific in terms of what... completion looks like. Like I actually, in my community, I gave feedback to the 200 people in my group on their websites.
And one of the biggest things I said is you have to understand what good realistically looks like. Now come down here. Let's say I want to do build.
Let's do one slice at a time. Honestly, just a nice little treasure trove of different skills that are really, really, really helpful based on the thing that you're doing. So the best way to use this is come over, see what you want to build.
What am I trying to do? test or ship a review, and you can just easily come down, grab this Lexo, copy it, and then you'll be ready to rock and roll. Now, here's the catch.
For plain, simple jobs, these are worse. They take longer and technically cost more of your credits because it's almost like sending a bazooka to open up a door, for example. It's just overkill.
It can be four times slower for very simple stuff. Then the other one I call, hey guys, in some circumstances, as you can see, using these skills, you will catch more, but I will be fair with you on this as well. This is honestly better for the cheaper models that you've got.
I think if you're using Fable 5 .2, it's a very exceptional model and increasingly will remove the need for these kinds of skills, so to speak. Honestly, it's more for the lower models where you're going to get the higher ROI. But bear in mind, it's going to take you more time to use this.
And so to use it, just simply add it in. Once you've done that, you can then type in the command. So spec, plan, build, test, review, or ship.
So realistically, this is something that's going to help you ship things in any stage that you're at. Probably one of the beneficial things here, guys, is just actually going through. this six -step process to make sure that you're actually building and launching things properly.
And then you can literally come down, guys, and just say, hey, I want you to work in this order, copy it and do that. And sometimes it's not even the skills themselves. It's just like, hey, this is a six -step process that we want to follow to make sure that our build and our plan is really good.
And honestly, just really handy if you think about it. For anything in this, you want to build an automated system for X. Well, we want to spec it out because we want to make sure that we cover everything, right?
You want it posting in the morning, in the evening. You want a daily improvement loop. We need to plan out what that looks like.
Then we need to build it. Then we need to... to test that that thing works then we need to review that the test was successful and then we do that wonderful thing at the end ship like a pirate so now we've covered the skills level five takes us nicely onto something called everything cloud code now cloud code gets it wrong okay it forgets and when it forgets it costs money now think of this as a building crew in the box now guys I did a full test.
I'm going to be brutally honest with what I found when I use this repo because it is blowing up and I want to show you exactly what the results are. Now in my testing, this is what I found. I compared it against plain code.
I used all 314 skills. which gives you just a lot of skills, guys. That's like a freaking construction site right there.
And then we've got one with just seven files only, okay? Now, obviously, as you can imagine, the fewer files we use, the cheaper it is. All of them scored 20 out of 20.
But of course, if you're loading in 314 skill, guys, oh my gosh, it's like double the price, which is crazy. Now, here's the thing. The tokens are higher, okay?
But if you use seven of them, the seven best ones, you're going to get essentially the same amount of tokens. as not using it without all of the bloat. But after testing them, guys, I'm going to recommend you take a couple of things.
Number one is the inspector. So the inspector itself checks your app for holes. It has an inspector checklist and also the inspector's house rules.
Two is a good one. It's called the diary. So essentially the diary itself writes down where you got to and it reads the diary back in a new chat.
Now, when I run this in Haiku, the inspector itself cost less, but it did take a little bit longer. And if you look at this, the diary actually worked. So there's one plain sentence, which is really cool.
If you want to go down and deep into this cloud code building crew in the box, I've actually got this full guide for you here, which talks about all 314 skills, the 71 different agents, how it all works, and essentially where cloud can kind of go wrong. And the issues it solves here, guys, is cloud can get it wrong. The way it solves that is it has an inspector with a checklist.
The second issue is it's time consuming. So what it does here, it has a site diary for the next. chat.
Okay. And finally, it's expensive. And what that means is small jobs actually go to a cheaper model, meaning that it just automatically routes it there for you.
And I'll put down below guys, all the results I got in terms of running all of them. So you can see exactly what you can get out of this pack of skills. And now we've covered everything called code.
We can now go on to level six, which is something called the slop monster. This one is incredibly powerful. And I can almost guarantee that you will find a use for this most of the week.
Now, the truth here is that AI copied sounds extremely predictable. like unbelievably predictable. What do I mean when I say that?
What the AI Slot Monster does is it goes through and takes any text, any copy, and what it does is it runs it through five clarity checks. So the first question is, what is it actually for? You give it any text.
It could be an email, it could be copy. And effectively, it has three different kind of like approaches to things. It can do for websites and what this has built into it, by the way, and this was built by basically looking at all of the best anti -slot repos on GitHub.
And what it effectively did is just combine the best elements of all of them. And I built... this one actually, guys, because I write copy all the time and I wanted something that works.
But where this takes it to a completely new level actually is the fact that it doesn't just do and remove the AI signs of the copy. It also actually does marketing and sales as well based on the very best kind of marketing and sales practices. But let me show you the difference there.
Look at this. And this is a before and after. I'm left on the side.
Look at this. Invoicing reimagined for freelancer. Send the invoice.
Northline chases the late ones. Look at this. Northline is the seamless invoicing platform that empowers freelancers to unlock a more effortless way.
to get paid. It's not just an invoicing tool. It's a game -changing journey towards financial freedom.
Oh my gosh. Get that AI off me right now. You see when a client opens your invoice, and if they still haven't paid, Northline sends a reminder so you don't write that awkward email yourself.
Guys, how ridiculous. And look at this. Can you see?
Everything that's highlighted is where it's identified these AI slop tropes. Unlock your potential. Start the free trial.
Look at this. What did it find? A number with no proof.
Not just X, it's Y. AI words listed free. Guys, I am telling you right now, this is one of the repos that I just...
Would not want to live without. It is really freaking powerful. And it's even shorted the words by like 40%.
It's freaking fantastic. Now let's check this one out. A cold email to a sponsor.
And it ranks the guys on AI sound and clarity. Check this out. Hi, Priya.
I hope this email finds you well. I wanted to reach out because I believe there is truly an exciting opportunity for Northline to leverage my channel, blah, blah, blah, blah. It's not just sponsorship.
It's a game -changing partnership. Look at this. Hi, Priya.
I'd like to make a sponsored segment for Northline in my desk lamp video on Thursday, the 10th of December. My viewers are freelance developers who send their own invoices so your product fits them. I'd use Northline.
offline. Keep telling me, blah, blah, blah, blah. Thanks, Theo.
Guys, I mean, like, are you picking what I'm dropping down here? This is crazy. Again, I reply to a refund request.
Just simple, clean, beautiful English. Like, genuinely, just absolutely fantastic. And this, guys, works on everything.
It works on emails. It works on websites. It works everywhere.
Now, to use it, you can clone it directly and just say, hey, go and grab the Sloth Monster. And as always, the link is down below in the pinned comment and the description. Now, the words read like a human now.
that's good okay however the page still looks like every other page which brings us on to asterix now asterix itself essentially here's the idea here i've got six prompts in front of us okay and six pages can you spot the pattern can you see guys what's jumping out in front of you the truth is all three made the same moves and actually guys ai design can be formulaic when you know what you're looking for.
A colored phrase in the headline. A second button, see how it works. A no card line underneath it.
So essentially, this is one of the fastest growing GitHub repos this week. And essentially, it is the box of parts Meta builds with. 174 components, 58 page templates, and across seven different themes.
And I compare them together. And I would say this, I'd say that honestly, it doesn't make better design, but it does remove... some of the AI slopification, okay?
What it is guilty of and what you would need to sanitize this with is the AI slop principles, okay? Because it has removed some of the AI slop, but at the same time, it does kind of also overwhelm you. Like, I wouldn't even know where to start looking on this page.
I would be like, where the hell do I even begin right now? Get me two extra coffees right now. Now, as you can see, look at this.
This is designed by Claude. I think we would all agree that this is a good design. However, it's freaking everywhere.
I could just see this cloud design a bazillion miles away. I could tell exactly who's designed them. This does not look like it was designed by cloud.
It would need a couple of follow -up prompts to say, hey, we need to focus on typography, spacing, that kind of thing. If you want a quick hack for this, check out the Power Design repo. It essentially codifies the 20 universal design principles, things like spacing, distances, and gaps, and things like that.
So again, this is essentially what we're dealing with. will remove the AI slarp, comma, but then it introduces a few other problems. And again, like, you know, meta is meta.
Like we've all used Facebook. Maybe you've used Facebook ads. You can kind of tell it's a good design.
It doesn't look like Claude's. If you do want to sell this to a client and you're like, hey, it doesn't look like it was vibe coded. It's pretty cool for that.
But I wouldn't say it's the end of the road. So now we've covered the anti -design sloppification. We now need to move on to the fact of like Claude's ability to build for us some pretty incredible games.
We have a GitHub guide there, guys, with a fantastic controller. Now with Claude, one prompt can make you a game.
That is awesome. But basically what this does here is this gives Claude a whole studio to build the game. So instead of saying, hey, make me a game.
It's a rough game. It's no brief, no check. With this, it asks what you want.
It writes a one -page brief. It builds a game and then checks a game against the brief. Again, just classic kind of...
what I would call looping methodology. Same idea, one prompt against the studio. And again, I've got these prompt games that we can play.
We've got Comma Hopper, which is great. You know, I joke sometimes saying like, actually, it's actually quite a cool game. I haven't played this.
It's pretty sick. Oh my gosh, look at this, guys. All right.
We've got this. It's pretty cool. I like it.
Let's come back and check this out. But it's a little bit different than the first one that we got, right? You can see like the level of detail, the level of kind of interactivity.
It's the first prompt. It isn't looking as good. This has six questions, five changes, but there is a kicker.
There's a fact that it was 20 times more expensive. The way that I would think about this is the fact that it will get a better quality, but what it does is just cut you out of it. Like this is not ready to go, obviously, right?
So you'd have to be the person sat down instructing it 10 times. This just kind of cuts out of the process. And I think that the quality of the guys, honestly, is a lot better.
Like you can tell this is really cool, right? This looks like a fun game, like probably something that you'd have bought on the app store, if I'm honest with you. Like it has that level of detail on it.
Again. First build, keep the planets big, make it fail, get the plants life. The catch is too quiet.
The flash rank is too heavy. It just goes through and kind of really helps you with the quality of that kind of thing. Okay.
Now a long prompt. If you were to give her, it would get close. But this is like a system.
It's essentially a skill, which is a system to do that. So if you want a prototype, you can just go in there and do slash prototype if you want to. But then you also have studio mode with 49 different agents.
You know, and here's just some examples of the games that people have built using this skill. I think it's so cool people can build and share these skills. Which leads us, very nice guys, onto level nine, which is all about heretic.
Now, what do I mean when I say heretic? And this is the truth, guys, is that these models are powerful, but they will refuse to do jobs. because they are censored models.
Now, the idea is this, is that you can ask a stock model a harmless request. I don't know, review this website or, you know, bolster my defenses for cybersecurity attacks. And they will say no, because like, and there've been examples, guys, the famous example now that happened with Hugging Face, they couldn't use frontier models to defend themselves.
They actually had to use GLM 5 .2 because the most powerful cloud models wouldn't do it because they were so hamstrung. And these models can be taken away from you anytime, which is why you want an uncensored model in your arsenal that will not refuse.
Now, what Heretic lets you do is take an open source model. And if you like uncensored, and what it does is it kind of asks a question. And the way that it works is.
When you ask it a question and it denies something, in kind of the multidimensional space, there's something that starts flagging up, which is like the bit that activates when it's trying to censor something. And it identifies where that is, and essentially it removes it. That's kind of the process of obliteration, if you like.
Guys, I actually tested this. Now, when I first ran this model here, which is the QAM 0 .5 billion parameters, a pretty small model, it refused 92 out of 100 of my controversial requests. After I obliterated it, it only refused four out of 100.
But the issue here is that's just scratching the surface of uncensored models. So the next thing we need to do is learn exactly why this is one of the most important things for you to understand. And we'll do that together inside this video right here.
The Hook

The bait, then the rug-pull.

The pitch is nine free repos that make Claude feel illegal. The useful part is the receipts: each tool is run against plain Claude Code with times, token counts, and dollar costs, and most of them come with a clear note on when you should not bother.

Frameworks

Named ideas worth stealing.

01:05model

Driver and car

  1. Model = the race car driver
  2. Harness = the car

Agent performance comes from two parts. Keeping the Claude Code harness while swapping in a free or local model keeps the skills and tools and only changes the brain.

Steal forexplaining why tool and workflow setup matters as much as model choice
18:13list

Six-stage build path

  1. Spec
  2. Plan
  3. Build
  4. Test
  5. Review
  6. Ship

Agent Skills organizes 25 skills into six stages, each with a slash command. Plan turns a spec into small ordered tasks, each with its own done test.

Steal forany automation or app build checklist
22:09list

Three Claude Code failure modes

  1. Gets it wrong: inspector with a checklist
  2. Time-consuming: a site diary for the next chat
  3. Expensive: route small jobs to a cheaper model

The useful core of a 314-skill pack, mapped to the problem each piece solves.

Steal fortrimming a bloated skill or agent setup
25:38list

Three AI landing-page moves

  1. A colored phrase in the headline
  2. A second 'see how it works' button
  3. A 'no card required' line underneath

Six AI-generated pages made the same three moves, which is why audiences call them slop even when they cannot say why.

Steal forauditing any AI-built landing page before it ships
04:15concept

Free model fit test

  1. Good: small jobs, trying Claude Code, chat and ideas, learning and demos, a spare brain
  2. Bad: speed, big jobs due today, walk-away tasks, private client work

When free routed models make sense inside Claude Code and when they do not.

Steal fordeciding which tasks to send to a cheaper model
CTA Breakdown

How they asked for the click.

VERBAL ASK
30:38next-video
“So the next thing we need to do is learn exactly why this is one of the most important things for you to understand. And we'll do that together inside this video right here.”

Ends mid-thought on uncensored models and hands off to a previous video, so the last repo doubles as a cliffhanger. Earlier asks: the free repo pack at 2:53 and the paid community at 11:23.

Storyboard

Visual structure at a glance.

open
hookopen00:00
free model demo
prooffree model demo01:59
22-run test
value22-run test04:09
agent desktop
valueagent desktop05:45
58,800 vs 2,900
proof58,800 vs 2,90009:43
community pitch
ctacommunity pitch12:21
cardboard
valuecardboard15:03
six stages
frameworksix stages18:21
314 vs 7
proof314 vs 721:31
slop before/after
proofslop before/after24:06
level grid
rehooklevel grid25:24
heretic
payoffheretic30:10
Frame Gallery

Visual moments.

One-click upgrade to your Google

Get more breakdowns in your search results

Add Modern Creator as a preferred source and Google shows you more of our breakdowns in Search, Top Stories, and AI Overviews. It only changes what you see, and you can undo it in your Google settings anytime.

Add to Preferred SourcesOpens your Google source preferences with us pre-loaded. Tick the box and you're done.
Watch next

More from this channel + related breakdowns.