The argument in one line.
Out of roughly 500,000 public Claude Code skills, only a handful survive daily use, and the ones that do each fix one specific, measurable failure: wasted tokens, over-trusted agents, or unexamined AI output.
Read if. Skip if.
READ IF YOU ARE…
- You already use Claude Code or Claude skills daily and want to prune your setup down to what actually earns its place.
- You run multiple AI agents in parallel and have been burned by output you didn't fully trust.
- You ship AI-generated code or copy and want a second AI pass built specifically to catch what the first one missed.
SKIP IF…
- You've never used Claude Code or an AI coding agent, most of this assumes an existing skills workflow.
- You're looking for a single silver-bullet tool rather than nine separate, narrow fixes.
The full version, fast.
A creator who tested hundreds of publicly available Claude Code skills ranks the nine he still uses every day. Caveman strips Claude's verbose preamble to save tokens. PStack argues you can't safely run parallel agents until you trust one rigorous agent first. VybeSecurity audits AI-written code for stack-specific vulnerabilities like hardcoded keys and client-trusted pricing. No AI Slop removes recognizable AI-writing tics without flattening the writer's voice. Ponytail and Matt Pocock's base-layer skills cut code volume and give agents a foundation to work from. Hyperframes generates motion graphics from spoken descriptions, and Taste plus Impeccable pair a one-shot design generator with a consistency auditor. The top pick, Auto Researcher, was built to train AI models by running hundreds of small experiments, and the video demonstrates that the same run-many-experiments method works on a personal gym routine just as well as on code.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 00:51
01 · Cold open — the case for skills
Argues most of the 500,000 public Claude skills are junk, but a few genuinely change how a model behaves once you've seen the pattern they fix.
00:51 – 01:56
02 · #9 — Caveman
The most-installed skill on the list strips Claude's preamble and hedging; creator claims a 65% token reduction, independent testing found closer to 35%.
01:56 – 03:30
03 · #8 — PStack / Potato Mode
Built by a React-compiler contributor now at Cursor: make one agent rigorous before running agents in parallel, because the real risk of multi-agent work is trust, not coordination.
03:30 – 05:09
04 · #7 — VybeSecurity
A stack-aware security audit built by Chris Roque (Aloha) that catches AI-written vulnerabilities like hardcoded API keys and client-trusted pricing before they ship.
05:09 – 06:27
05 · Sponsor: HubSpot's free Claude workflow guide
Mid-roll for a free guide mapping Anthropic's product lineup, model picks, and three example workflows.
06:27 – 08:35
06 · #6 — No AI Slop
Built by Peter Yang to strip 22 recognizable AI-writing patterns without sanding off the writer's actual voice, and it logs every change so you can argue with it.
08:35 – 09:45
07 · #5 — Ponytail
Makes Claude write like a terse senior engineer; a claimed 94% reduction in code volume measured closer to 50% in the creator's own test, plus cheaper and faster runs.
09:45 – 11:19
08 · #4 — Matt Pocock skills
The base-layer skill set the creator now runs under everything else, replacing Superpowers; includes a code-review skill used to make one model grade another's output before it reaches him.
11:19 – 12:51
09 · #3 — Hyperframes
Turns a spoken description into rendered motion graphics; an eight-minute firework animation the creator argues beats a full day in After Effects.
12:51 – 14:31
10 · #2 — Taste + Impeccable
Two design skills that survive together: Taste one-shots a design direction on a blank page, Impeccable then audits a real multi-screen app for the inconsistencies nobody looks at twice.
14:31 – 16:07
11 · Honourable mentions
Four skills that almost made the cut: Unlazy (anti-laziness via a depth-tree method), AI Job Search, Agent Reach, and Open Design, an open-source alternative to Claude's Core Design.
16:07 – 18:22
12 · #1 — Auto Researcher
Originally built to train AI models by running hundreds of tiny experiments and keeping what works; the creator argues the method, not the machine-learning use case, is the real value.
18:22 – 21:18
13 · Putting #1 to an unexpected test
Runs the Auto Researcher methodology on a personal five-day gym split instead of code, producing a restructured program in minutes to prove the skill generalizes past coding.
Lines worth screenshotting.
- A tool's own efficiency claim is a starting estimate, not a fact: one skill claimed a 65% token reduction and independent testing measured closer to 35%.
- The failure mode of running many parallel AI agents isn't coordination, it's trust: you can't safely scale out what you haven't first made rigorous.
- A browser that lets the client decide a product's price means anyone can open DevTools, change the number, and have the payment processor accept it.
- AI didn't just write insecure code by accident, a person then confidently approved that insecure code without catching it.
- Using AI to polish AI-written text tends to erase the exact irregularities, odd sentence lengths, specific analogies, that made it sound human to begin with.
- A code-volume skill claimed a 94% reduction in output; independent testing measured roughly 50%, still enough to also cut cost and runtime.
- Having one model review another model's output before a human sees it catches mistakes inside the loop instead of after the fact.
- Motion graphics have one of the worst effort-to-payoff ratios in content production, a four-second animation can cost an hour of manual keyframing.
- Design inconsistencies, like four grays that should be one, are the kind of thing you look directly at without seeing, which is what a dedicated audit pass exists to catch.
- A tool built to train AI models by running hundreds of tiny experiments is really just the scientific method with the patience problem removed.
- You're not limited by what an experimentation tool can do, you're limited by what you thought to point it at.
- The same run-many-experiments methodology that improves a codebase restructured a personal gym program in minutes, proving the method, not the domain, is the value.
Which Claude skills survive daily use, and why.
WHAT TO LEARN
Out of hundreds of tested Claude Code skills, the ones that stick each address one specific, measurable failure, wasted tokens, over-trusted agents, unexamined AI writing, and a scientific method for improving anything you point it at.
02#9 — Caveman
- Stripping a chatbot's throat-clearing preamble isn't just about tone, it measurably cuts the tokens you pay for on every response.
- Independently verifying a tool's own performance claim is worth doing before you build a workflow around a number someone else published.
03#8 — PStack / Potato Mode
- Running many AI agents in parallel doesn't fail because of coordination, it fails because you don't yet trust what any single agent produces.
- Making one process rigorous before you scale it out catches problems that only show up once you've felt the difference between careless and careful output.
04#7 — VybeSecurity
- AI-written code introduces vulnerabilities like hardcoded API keys and client-side price checks with total confidence, and confident code still needs a second, adversarial pass.
- A security check that understands your specific stack catches real issues that a generic 'check my app for problems' prompt misses.
- If a browser can decide how much a product costs, an attacker can set that price to whatever they want, so trust never belongs on the client.
06#6 — No AI Slop
- AI writing has a recognizable set of tells, and once you can name the pattern you can't unsee it.
- Sanding AI phrasing off AI text with more AI often erases the specific, uneven details that made the original sound human in the first place.
- A tool that lists every change it made, rather than silently rewriting, lets you argue with an edit instead of accepting it blind.
07#5 — Ponytail
- The core complaint about AI-generated code usually isn't quality, it's volume, so a tool optimized purely for writing less code addresses a real, separate problem from correctness.
- Treat a vendor's efficiency claim as a starting estimate, not a fact.
08#4 — Matt Pocock skills
- A general-purpose base layer of skills works when the tool's own descriptions are specific enough for it to self-select correctly, without hand-picking per task.
- Having one model review another model's output before a human sees it catches mistakes inside the loop instead of after the fact.
09#3 — Hyperframes
- Motion graphics have one of the worst effort-to-payoff ratios in content production, which makes natural-language generation worth testing even where existing tools already work.
- A single spoken description turned into a finished, particle-heavy animation in minutes shows how much manual keyframe work was pure friction, not creative judgment.
10#2 — Taste + Impeccable
- A blank page and an already-shipped product need different tools: one that invents a point of view fast, and one that grinds out consistency after the fact.
- Design inconsistencies are the kind of thing you look directly at and don't see, which is exactly what a dedicated audit pass is for.
11Honourable mentions
- A depth-tree method that gives every sub-task the full time budget of the whole task directly counters a model quietly doing most of the work and calling it done.
- The most immediately valuable tool on a list isn't always the flashiest one.
12#1 — Auto Researcher
- A tool built to train AI models by running hundreds of experiments and keeping what works is really just the scientific method with the patience problem removed, and that generalizes far past machine learning.
- You're not limited by what the tool can do, you're limited by what you thought to point it at.
13Putting #1 to an unexpected test
- Running the same experiment methodology on a personal routine instead of code proves the tool's value is the process, not the domain.
- A no-cost, no-hardware experiment can restructure a real decision in minutes instead of weeks of trial and error.
Terms worth knowing.
- Claude skill
- A packaged instruction file (a SKILL.md plus supporting scripts) that changes how Claude Code behaves for a specific task, installed on top of the base model.
- Token
- The unit AI providers bill by; shorter, more direct model responses use fewer tokens and cost less per request.
- Superpowers
- An earlier popular Claude skill bundle that this video argues has been surpassed as the foundational base-layer skill set.
- Core Design
- Anthropic's built-in design tool for Claude, which cannot be extended with third-party skills, the gap an open-source alternative is built to fill.
- Depth-tree method
- A task-splitting approach that gives every sub-task the full time budget of the whole task, so effort doesn't thin out the deeper an agent goes.
Things they pointed at.
01:03toolCaveman
02:20toolPStack / Potato Mode
03:45toolVybeSecurity
06:31toolNo AI Slop
09:05toolPonytail
10:30toolMatt Pocock skills
11:19toolHyperframes
13:10toolTaste
13:20toolImpeccable
14:45toolUnlazy
15:00toolAI Job Search
15:10toolAgent Reach
15:40toolOpen Design
16:29toolAuto Researcher
Lines you could clip.
01:03
“Why use many token when few token do trick?”
03:02
“The problem is not running parallel agents. The problem is trust.”
03:45
“Someone opens DevTools, changes your $99 product to 1 cent, hits buy, and Stripe just processes it.”
06:45
“When you use AI to fix AI writing, it polishes out the exact stuff that made it sound human.”
09:05
“The problem is not that Claude writes bad code. It actually writes pretty good code. It just writes far too much of it.”
16:29
“Run hundreds of tiny experiments, keep what works and then throw away what doesn't. Repeat and that's it.”
20:50
“Everything else on this list makes you faster at something you were already doing. This one actually changes the size of the problem you're willing to attack.”
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
There are over 500 ,000 cloud skills out there, GitHub is full of them. But most of them don't even work, and some are just nicely formatted system props with a readme file. Very few are actually worth installing, and you might think you don't need them.
Models are already powerful enough, right? Why add a skill file? But here's the thing, every model follows its own patterns, and once you've seen it, you can't unsee it.
That's why you need something to steer it. These skills are built by people in the trenches. You've got a meta engineer, a security researcher, and they've repackaged what works for them so you can use it too.
From a skill that cuts your target usage by 65 % to one that creates motion graphics using natural language. I've tested hundreds of skills at this point, and these nine are the ones I keep coming back to every single day. Thanks to HubSpot for sponsoring this video.
We're starting with the absolute holy grail of clawed skills. The skill that arguably started this whole craze with over half a million installs. Of course, I'm talking about the caveman skill because Why use many token when few token do trick?
If you haven't heard about the caveman skill before, then chances are that you are probably a caveman yourself because this is unironically one of the most useful skills I've ever installed and probably the most entertaining as well. We all know Chord is a complete yapper. You ask for one simple request and you get three paragraphs of great question.
Let me walk you through my approach. That preamble is not only annoying, but it costs few tokens. tokens equals money what i love about caveman is it drops all of it and just gives it to you straight it's like having a spam filter for your inbox but it also saves you tokens the creator claims a 65 reduction in token saves but i actually tested this myself and it was closer to about 35%.
I'd rank caveman higher, but everyone already knows about it. So I'll leave it at number nine. Next is a new skill that's been gaining attention from an up and coming developer.
Millions of lines of code at meta, Netflix, and now cursor. The person who helps build and maintain the React compiler, that is who wrote number eight. Of course, I'm talking about Pstack.
Pstack is a set of skills from Potato with an E, and the one that everyone's going crazy about is Potato Mode. Her pitch is one of my favorite lines I've ever read in a repo. I don't want to ship like a team of 20 slob artists.
And then she says, if you want to go fast, you first need to go deep. which is the exact opposite of how everyone is currently using agents, by the way. The move right now is spin up eight parallel agents and pray.
And you can do it. It technically works. It feels incredible for about 20 minutes until Sunday comes around and you're looking at 8 ,000 lines of code trying to figure out which one of your robot children broke your app.
The insight in PStack is that the problem is not running parallel agents. The problem is trust. you cannot safely run 10 agents if you do not trust what the agent produces make one agent rigorous first then run parallel agents once you have felt the difference you cannot go back once you go deep you can't go to sleep and the cherry on top is the fact that these are the actual skills she uses every day at cursor to shift thousands of lines of code every single day okay okay number eight was pretty solid but if you thought that was good Wait till I tell you about rank 7.
Hardcoded API keys, security rules completely skipped, auto tokens sitting in local storage, and my personal favorite, letting the browser decide how much something costs. Do you know what that means? Someone opens DevTools, changes your $99 product to 1 cent, hits buy, and Stripe just processes it, because your app basically asks the customer How much would you like to pay?
So number seven is VybSecurity built by Chris Roque. I never know how to pronounce his name and the team at Aloha. If you've been anywhere near this side of YouTube, you probably already know Chris.
I watch him all the time. And this isn't just some generic prompt that says, check my app for security problems. It actually understands what your app is using.
If you're using SuperBase, it checks for SuperBase specific security issues. If you're using Next .js, it checks for Next .js issues. And I think this skill is massively underrated.
It has under a thousand stars. Because the scary part about bytecoding isn't that you intentionally wrote insecure code. The AI did.
It confidently hardcoded your API key. It confidently skipped your security rules. and you confidently clicked approve on it so use ai to check ai run this on your app every time you make a big update i ran around a side project i'd already shipped six weeks earlier and it found a few fatal problems all right vibe security catches the vulnerabilities you did not know claude wrote but there is another blind spot almost everyone has and it is bigger than security so here's an uncomfortable truth i learned while researching for this video Most of you are using maybe 10 % of what Claude can actually do.
You open the chat, you type, you close it. That is the whole workflow. Meanwhile, skills, co -work, code, and design are just sitting there, uncut.
And I get it. Anthropic ships new stuff faster than anyone can track. It is confusing.
That's why I wanted to share this free guide called flawed at work from hubspot that maps the entire ecosystem it has every product side by side so you can see where each one fits it even has the current model lineup with which model should i use breakdown and three real world workflows showing you how to chain these tools together this guide will show you exactly how the skills like the ones we're talking about in this video can fit into your daily workflow in a way that actually multiplies their capabilities My favorite one is the workflow that takes a single video and turns it into a full week of content across platforms, step by step with which cloud product to use at which stage.
If you're someone who makes content, that one workflow alone is worth the download. It takes one read and it will show you capabilities you did not know you had. Check it out in the description.
And thanks to HubSpot for sponsoring today's video. All right, back to the skills. Because number six is about fixing the thing Ford is genuinely the worst at.
No AI slop. Built by Peter Yang, it kills 22 patterns of AI writing that have infected basically everything on the internet right now. You probably already know these.
A lot of us have developed a physical reaction to them. It's not X, it's Y. The future isn't coming, it's already here.
What nobody tells you is, and don't even get me started on the em dashes. Every single platform, LinkedIn, Twitter, Substack, YouTube comments, they are everywhere. And here is the trap.
When you use AI to fix AI writing, it polishes out the exact stuff that made it sound human. Your weird sentence length, your bad analogies, the joke that only lands if people know you. Those were the good bits.
Most writing skills sand them off and hand you back a nicer flavor of the same problem. This skill's actual trick is it removes the patterns without flattening your voice. And then this is the part I genuinely love.
It lists everything it changed so you can actually argue with it. All right. So as you can see, I've made Claude write a short sales script.
about my app ace and it is full of ai terminology here's the part that matters it's not x it's y wow so we're gonna go ahead and run the no ai slop skill let's see how it does and boom as you can see we now have a script and ace is a one -time payment $150 for lifetime access, including every future update Ace ships, no recurring fee ever.
That's a huge difference from, here's the part that matters. Ace isn't a subscription, it's a one payment. Like that's a way better improvement.
And here you can see what actually changes. So it's really easy to understand, rewrote the offer line. And then if you want, you can argue in and be like, hey, I actually didn't want you to rewrite the offer line.
It's really cool. That is the writing slot dealt with. Number five deals with the code slot problem.
Coming in at number five, and of course, I'm talking about Ponytail. And Ponytail makes your agents think like . He says nothing, he writes one line, and it works.
You know the guy, every team has one. You spend two days building an app, you're showing them, and he looks at it for eight seconds and goes, why didn't you just do this? This skill is the embodiment of that ponytailed individual.
A perfect anti -slop skill for code volume. Because here is the thing a lot of people don't realize about AI coding. The problem is not that Claude writes bad code.
It actually writes pretty good code if you know what you're doing. It just writes far too much of it. Ponytail claims to reduce the amount of code Claude writes by up to 94%.
But from my own test, it's closer to around 50%. And there's some pretty neat side effects too, like pass being about 20 % cheaper and 27 % faster. Everything is verifiable on the repo, but I recommend you try it for yourself.
Ponytail is one idea executed perfectly. Write less code. Number four is the entire operating system.
And I'm about to say something that might ruffle a few feathers. Remember superpowers? Everyone lost their minds over it.
every thread, every video, every this changes everything post and honestly it deserved it at the time it was good but this replaces it. I'm talking about Matt Pocock skills and I'm not gonna lie it is the best base layer I have ever used because this is not a clever one trick skill like caveman or ponytail it is the foundation you put underneath everything else.
Here's how I actually use it and if you steal one thing from this video Still this, I give my agent a link to the repo, kick off a long -running task, and tell it to use the skills where they make sense. That is the whole instruction.
I do not micromanage which skill fires where because the descriptions inside the skills are actually good enough. It just picks correctly. For example, inside Matt Pocock's skills, there is a code review skill.
I have Claude Fabel orchestrate Sonnet models to do the heavy execution work. It then uses the code review skill to make Sonnet review its own output before it ever comes back to Fabel, the orchestrator agent. The swap actually gets caught inside that loop.
The same model that wrote it is now the same model that judges it. It never hits my eyeballs. I just get a finished product that's ready to test.
Alright, foundation, reviewer, security, editor, all the boring stuff is sorted, which means we can finally get to the fun stuff. Alright, what if I told you every single piece of motion graphics in this video that you're watching was made using the skill hyperframes? And I'm going to say this straight up, but I think it's actually better than Remotion.
Remotion's great, but I'm not trying to start a fight. Hyperframes just produces better outputs more of the time with less nagging. So if you make content, you already know the problem.
Motion graphics are the single worst effort to pay off ratio in the entire pipeline. A four second animation that makes one point clearly can eat an hour of your life in After Effects. Then you watch it back and go, actually, the easing is wrong.
So here's what I'm going to do. I'm going to describe an animation out loud right now and we're going to see what comes back. I'm going to link the hyperframe skill and just say create an animation of a really big crazy firework explosion in the night sky.
And I'm just gonna let this run and let's see what it generates. And we're done. That took roughly eight minutes.
Let's have a look at the video.
Damn, that looks nice. Wow, I love how it explodes and shakes the sky. Well, there you go.
In eight minutes, we were able to create this sick ass motion graphics. Imagine trying to do that in After Effects. That would literally take you an entire day to get all the particles to look like that.
Hyperframes makes cool animations, but coming in at number two, it turns Claude into an award -winning designer. All right, this one's actually cheating slightly, but coming in at number two, it's actually two skills. No, no, no, you can't do that, no.
Hey, it's my video. I make the rules around here. But I'm not gonna lie, there are a lot of design skills out there.
Genuinely, dozens at this point. And I'm going to say something that sounds harsh, but I mean it. most of them are like the same six prompts about white space wearing different hats these two are the ones that survived on my machine and the reason they both survived is they are good at completely opposite things taste is the one -shot king if it's a blank page no existing design language you need something that looks intentional on the first try taste just hands it to you it is the i need a landing page in 20 minutes and i need an asap kind of skill impeccable is the opposite it's the analyst it's not trying to be clever it is trying to be right put out a real app with 15 screens and it goes through them properly one by one maybe you've got inconsistent spacing you've got four different grays that should be one gray that button that's like two pixels off on one screen only things you have absolutely looked directly at and not seen and the combination of these two is where it gets stupid
Taste for the first pass so you have something with a point of view. Impeccable to grade it and grind out every inconsistency. One gives you the idea, the other gives you the discipline.
I've shipped things with this combination that people assumed I paid a designer for. Actually, no. Before number one, here are four skills that did not make the list but absolutely could have because I've seen comments about every single one of these and I would rather address them now than read about it in the comments for the next six months straight.
First, unlazy. This one genuinely hurt to cut out. It is an anti -laziness skill for agents built around something called the depth tree method.
Here's what it does. It splits a task into layers and gives every single leaf the full time budget of the whole task. So effort actually multiplies with depth instead of getting thinner.
It's built on actual research into model laziness and premature completion. That thing Claude does where it writes about 80 % of the code and then types the rest follows the same pattern. I hate that.
Unlazy fixes it. So if you do long agentic runs, this is amazing. Go get it.
Next, you've got AI job search. Exactly what it sounds like. If you are job hunting right now, this is probably the most immediately valuable thing in this entire video.
And it did not even make the list. Next, you've got agent reach. Give your agents access to gated parts of the internet.
Use it responsibly though, please.
Fourth, you've got Open Design and this one fixes a genuinely annoying limitation. Core Design is one of my favourite creations from the Anthropic team, but you cannot extend it with skills. OpenDesign lets you.
It's basically an open sourced free core design alternative that lets you use skills like impeccable and taste to improve on your UI screens. And I want to reframe something real quick because number one is not the flashiest skill on this list. Hyperframes makes pretty things.
Ponytail has better numbers. Number one is here for a different reason. It is probably the most versatile, most underrated and biggest upgrade to your agent skill artillery today.
Now this skill was originally built to train AI models. The mechanism is brutally simple. Run hundreds of tiny experiments, keep what works and then throw away what doesn't.
Repeat and that's it. That is the whole thing. Now read that again and tell me what part of it is actually about machine learning.
Nothing. Nothing about that is specific to models, it is just the scientific method. With the patience problem removed, because the reason you and I do not run 100 experiments on anything in our lives is not that we do not believe in the method, it's that 100 experiments is 100 experiments.
So run 100 experiments to optimize your budget. Run 100 experiments on your calorie intake. Run 100 experiments on your app's UI.
Run hundreds of experiments on your thumbnails. You are not limited by the skill. You are limited by what you thought to point it at.
All right, now a lot of people don't know this, but when I first built Ace, I actually ran the auto researcher skill on my app and it was able to create a lot of tiny little delight moments within my app. Like if we just click on any of these and we look at the whiteboard, even the loading screen, you can see like just being able to connect the models and stuff like that it created a lot of tiny delight moments like literally hundreds of them and they stacked up on top of each other to the point where like i don't even know what is and what isn't done by the auto research skill anymore things like the line glowing when it's connected the ability to change the things the pill of the model picker here like how that menu looks a lot of this was designed and implemented by the auto researcher skill but we're not going to be doing that today for today's example i've actually got my gym workout and i've pasted it in it's a five day upper lower split all i'm gonna say is use the auto research skill methodology because we're not using the skill because it's used to train ai agents we're using the methodology of the skill to run a hundred tiny experiments to improve my gym routine i want to gain hypertrophy around the arms look bulkier and tone i'm 167 176 centimeters tall
And I weigh 80 kg. I'm a 25 male. Now I'm going to go ahead and send that front.
And I'm doing this just to prove that you can literally run this skill on anything and it will find the most optimal best solution. Just like when I ran it on my app and I focused it on finding, you know, animations and tiny little delight moments that would make a user want to stay on the app and have a more enjoyable experience.
It's the exact same thing I'm doing with my workout routine. And I'm using Opus 5 on medium effort because you need to also consider that running 100 experiments is not cheap. So I'm going to let this run and let's see how it does.
Okay, so as you can see, it went ahead and ran 100 experiments. Base diagnosis, five things wrong in effect order. Direct arm volume is seven sets a week each.
Priority range is 12 to 20. Five of seven triceps sets are push downs, no stretch position biceps work at all. So yeah, pretty interesting.
Let's have a look and boom. If you compare it to the previous plan, you can see day one, we've got barbell bench press, three sets, six to eight reps, lat pull down, inclined bench, seated cable row. This all seems to look the same.
Okay, except we changed it to an overhead cable easy tricep extension And then I also did incline dumbbell curls. Okay, so it's changed that slightly Day two we've got lower Which is unchanged and then day three it's actually doing arm priority So it's changed this completely to weighted dips or close grip bench weighted chin -ups overhead press chest supported.
So very arm heavy movements here lower is the same then upper it's actually added a bit more of what we missed in day three super cool and there you go everything else on this list makes you faster at something you were already doing this one actually changes the size of the problem you're willing to attack and it costs nothing it is just a text file And there you have it, the secrets to all my favorite skills.
I've left a link to all of them in the description. Leave a like on your way up. Peace.
The Hook
The bait, then the rug-pull.
Half a million Claude skills exist on GitHub and almost none of them are worth installing. This is the short list of nine that survived one creator's daily use, plus the specific, measurable problem each one fixes.
Named ideas worth stealing.
01:03concept
Caveman
Strips Claude's hedging preamble ('let me walk you through my approach') so responses arrive direct, saving tokens on every reply.
Steal forany workflow where verbose AI preamble is wasting tokens or reading time
02:20concept
PStack / Potato Mode
Make one agent rigorous and trustworthy before running multiple agents in parallel; the constraint on multi-agent work is trust, not raw parallelism.
Steal forany multi-agent coding setup before scaling to parallel runs
03:45concept
VybeSecurity
A stack-aware security scan (Supabase-specific, Next.js-specific) that catches hardcoded keys, skipped auth checks, and client-trusted pricing that AI coding tools introduce with full confidence.
Steal forauditing any AI-generated app before it takes real payments
06:31list
No AI Slop — 22 patterns
- it's not X, it's Y
- the future isn't coming, it's already here
- what nobody tells you is
- em dashes
Removes 22 recognizable AI-writing tics while listing every change made, so the writer can argue with any specific edit.
Steal forcleaning up AI-drafted marketing copy without losing the writer's voice
09:05concept
Ponytail
Prompts Claude to write like a terse senior developer, cutting code volume by roughly half in independent testing versus a 94% vendor claim.
Steal forreducing AI-generated code bloat before review
10:30model
Matt Pocock skills
A general base-layer skill set applied broadly rather than hand-picked per task, including a built-in code-review skill for model-reviews-model checks.
Steal fora default skill foundation to run under every other skill
13:10model
Taste + Impeccable
- Taste: one-shot a design direction on a blank page
- Impeccable: audit a shipped app screen-by-screen for inconsistency
Pairs a first-pass design generator with a second-pass consistency auditor, covering both the blank-page problem and the already-shipped-but-inconsistent problem.
Steal forany UI that needs both a point of view and a polish pass
16:29model
Auto Researcher
- Run many small experiments
- Measure the result
- Keep what works
- Discard what doesn't
- Repeat
Originally built to train AI models; applying its run-hundreds-of-experiments method to any decision (app UI, a workout plan) removes the patience problem that stops most people from ever running that many trials.
Steal foroptimizing anything you'd normally guess at instead of test
How they asked for the click.
VERBAL ASK
05:09link
“Check it out in the description. And thanks to HubSpot for sponsoring today's video.”
Standard mid-roll: pitch a free workflow guide, then bridge straight back into the countdown.
MENTIONED ON CAMERA
Visual structure at a glance.
