19:19Dubibubi · Review
I Made Claude Opus 5 and Fable 5 Build the Same App
A blind, three-build test of Anthropic's newest coding models — and the first time this reviewer picked a model other than Fable as the winner.
July 25thThe argument in one line.
Out of 85,000 community-built Claude Skills, install data shows only a small, repeatable set actually earns its place — three of them (Context7, Superpowers, Frontend Design) belong on every project, and the rest should be picked to match your stack, not collected wholesale.
Read if. Skip if.
READ IF YOU ARE…
- You already use Claude Code or a similar agentic coding tool and want to know which community skills are worth installing.
- You're a solo or small-team builder without a dedicated security or QA function and want AI-assisted guardrails.
- You're curious which agent-skill categories (language servers, testing, review, design) actually move the needle on output quality.
SKIP IF…
- You don't use Claude Code, Cursor, or another agent-skill-compatible tool — the specific installs won't apply.
- You're looking for a tutorial on writing your own skill from scratch rather than a ranked buying guide.
The full version, fast.
The creator scraped install-count data across 85,000 community Claude Skills and ranked the 12 most installed, from Security Guidance (175K installs) up to Frontend Design (829K installs, the single most-installed skill on the platform). Along the way he covers language-server skills (TypeScript LSP, Pyright), full agentic workflows (Feature Dev, Code Review), memory/context tools (CLAUDE.md Management, Context7), the GitHub MCP connector, Skill Creator, Playwright browser testing, Code Simplifier, and Superpowers — a full test-driven-development framework. Three honorable mentions (Caveman, G-Stack, Karpathy Skills) round it out. The closing rule: don't install all 12 — three to five skills matched to your actual workflow is the sweet spot, because every skill eats context window and adds slash commands, and roughly one in eight community skills is flagged as containing a security issue.
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 – 01:27
01 · Cold open + promise
Leads with the #1 skill's install count, frames the 85,000-skill glut as mostly noise, promises a countdown from 12 to 1 plus a rule for not over-installing, and introduces himself ($40K in 50 days using these tools).
01:27 – 02:48
02 · 12. Security Guidance
175K installs. Checks every Claude file edit, commit, and move against OWASP-style security frameworks before it runs; cites a stat that ~1 in 8 community skills contain security issues; recommends pairing with Semgrep's free security skills.
02:48 – 03:36
03 · 11. TypeScript LSP
177K installs. Gives Claude a live map of types, references, and imports instead of reading files blind, cutting hallucinations; notes the same language-server pattern exists per-language (Pyright for Python at 91K installs, plus Go, Rust, Java).
03:36 – 04:08
04 · 10. Feature Dev
217K installs. A five-step agentic pipeline — explore codebase, design architecture, write code, generate tests, review — that runs a small dev team in sequence instead of one pass.
04:08 – 04:43
05 · 9. CLAUDE.md Management
223K installs. Keeps the project's CLAUDE.md memory file updated automatically as the project grows, so conventions and stack details don't have to be re-explained every session.
04:43 – 05:28
06 · 8. Playwright
248K installs. Gives Claude an autonomous headless browser to click through the UI, verify features work, and fix broken tests; notes Claude's own creator built a custom /go skill around this.
05:28 – 05:55
07 · 7. GitHub MCP
261K installs. The most-installed MCP connector overall (ahead of Vercel, Figma, Slack); plugs Claude directly into GitHub so it can act on the repo without leaving the session.
05:55 – 06:45
08 · 6. Skill Creator
283K installs, official Anthropic-backed. A skill that builds other skills — describe a task in plain English and it drafts, tests, and packages a new reusable skill.
06:45 – 07:10
09 · 5. Code Simplifier
284K installs. Runs a compression pass on AI-generated code — strips redundancy, simplifies logic, cuts technical debt without changing behavior.
07:10 – 07:42
10 · 4. Code Review
347K installs. Automates code review end to end: checks changes against standards, writes PR descriptions, flags architecture problems; often paired with a PR review toolkit that adds confidence scores.
07:42 – 08:18
11 · 3. Context7
348K installs. Live documentation lookup and smart search across large codebases so Claude pulls exact docs/patterns on demand instead of only knowing what fits in context — one of three skills recommended for every project.
08:18 – 09:03
12 · 2. Superpowers
752K installs, an entire framework rather than a single skill. Forces brainstorming before coding, breaks work into atomic tasks, and enforces test-driven development — restarts if Claude writes code before a failing test.
09:03 – 09:52
13 · Sponsor break: Ace
Mid-video ad for Ace, an app for orchestrating multiple AI subscriptions (Claude, Codex) as a coordinated team with loops and delegation; discount code given.
09:52 – 12:15
14 · Honorable mentions
Three skills that didn't make the top 12 but are called out anyway: Caveman (strips Claude's verbose explanations, claimed ~75% output-token reduction), G-Stack (Garry Tan's full startup-workflow Claude setup: CEO/eng/design/QA/security review stages), and Karpathy Skills (four rules distilled from Andrej Karpathy's complaints about AI coding, packaged as a single CLAUDE.md).
12:15 – 13:25
15 · 1. Frontend Design
829K installs, the most-installed skill on the platform. Forces Claude to commit to one bold, distinctive visual style before writing UI, fighting generic 'AI slop' aesthetics; named as the third leg of a recommended global trio alongside Context7 and Superpowers.
13:25 – 14:26
16 · The security warning + the one rule
Anthropic itself warns skills run code with your permissions and some have been documented hijacking tool routing; rule of thumb is only install from the official marketplace. Closing rule: don't install all 12 — three to five skills matched to your actual workflow is the sweet spot, since every skill eats context window and adds slash commands.
14:26 – 14:46
17 · Outro
Points to a related video and signs off.
Lines worth screenshotting.
- Front-end Design is the single most-installed Claude Skill on the platform at 829,000 installs, roughly 25% more than the #2 skill.
- A security audit found that of 85,000 community Claude Skills, about 11,000 — one in eight — are believed to contain security issues.
- Superpowers, the #2 skill at 752,000 installs, forces Claude to brainstorm and write a failing test before writing any code, restarting if that order is broken.
- TypeScript LSP gives Claude a live map of every type, reference, and import in a codebase instead of making it read files blind, which directly cuts hallucinations.
- Skill Creator is a skill that builds other skills: describe a task in plain English and it drafts, tests, and packages a new reusable skill.
- The three skills recommended for global installation across every project are Context7, Superpowers, and Frontend Design.
- Claude Skills run on an open standard (agentskills.io), so the same skill installed for Claude also works in Cursor and GitHub Copilot.
- The sweet spot for total skills installed is three to five matched to your workflow — every additional skill consumes context window and adds slash commands.
- The GitHub MCP connector is the most-installed MCP connector of any of them, ahead of Vercel, Figma, and Slack.
- Code Simplifier runs a compression pass on AI-generated code specifically to strip redundancy and cut technical debt without changing behavior.
- The Karpathy Skills file compresses Andrej Karpathy's coding complaints into four rules — think before coding, keep it simple, make surgical changes, define success first — and became one of the fastest-growing single-file GitHub repos of 2026.
Three skills belong everywhere; the rest should match your stack.
WHAT TO LEARN
Install data across 85,000 community Claude Skills points to a small, repeatable set that earns its place — a language-server skill for whatever you build in, an infrastructure trio for every project, and a hard cap of three to five total.
0212. Security Guidance
- About one in eight community-built AI skills is believed to contain a security issue, so treat every skill install like a dependency you'd vet, not a free upgrade.
0311. TypeScript LSP
- Install a language-server skill (TypeScript LSP, Pyright, etc.) for your primary language first — it maps your codebase's types and references so the agent stops guessing at structure.
0410. Feature Dev
- A five-step pipeline (explore, design, write, test, review) turns a single agent pass into something closer to a small team working in sequence on one feature.
059. CLAUDE.md Management
- A memory-management skill that keeps your project's instructions file current removes the need to re-explain conventions every session.
068. Playwright
- Giving an agent an autonomous headless browser lets it verify its own UI changes and fix broken tests without you driving the browser yourself.
077. GitHub MCP
- The most-installed connector category isn't a coding skill at all — it's a bridge into an outside service (GitHub), which turns a local assistant into something that can act on your actual repo.
086. Skill Creator
- A skill that builds other skills means you're not limited to whatever's pre-packaged — describing a repeated task in plain English can become a reusable tool.
095. Code Simplifier
- AI-generated code tends toward bloat by default; a dedicated compression pass that strips redundancy without changing behavior is worth running periodically.
104. Code Review
- Automating code review against your own standards catches issues before a human looks and produces the PR description as a byproduct.
113. Context7
- Live, on-demand documentation lookup beats relying on whatever fits in context — it's the difference between an agent following your conventions and inventing its own.
122. Superpowers
- Forcing a failing test before any code is written, and restarting when that order breaks, is a concrete way to stop an agent from jumping ahead of your intent.
151. Frontend Design
- Committing to one distinctive visual style before writing UI code is what separates a distinctive interface from generic, interchangeable AI output.
- Because these skills run on an open standard, a workflow built once travels across tools rather than locking you into a single vendor.
16The security warning + the one rule
- The failure mode isn't under-installing, it's over-installing: every added skill costs context window and adds commands, so three to five matched to your actual work beats collecting all twelve.
Terms worth knowing.
- Claude Skill
- A packaged, reusable capability installed into an AI coding agent with a single command — bundles instructions, hooks, and MCP servers into a shareable unit that works across sessions.
- LSP (Language Server Protocol)
- A protocol that gives a coding tool a live map of a codebase's types, references, and imports, so the AI doesn't have to infer structure by reading files one at a time.
- MCP (Model Context Protocol)
- An open protocol that bridges an AI agent to a live outside service (like GitHub), letting the agent take real actions in that service without leaving the coding session.
- Superpowers
- A community framework that enforces a disciplined development methodology on an AI agent — brainstorm first, break work into small tasks, and require a failing test before any code is written.
- Test-driven development (TDD)
- A workflow where a failing test is written before the code that makes it pass, used here to stop an AI agent from jumping straight into unverified code.
- OWASP
- A widely used industry framework of common security vulnerability categories, referenced here as the standard the Security Guidance skill checks code changes against.
- agentskills.io
- The open standard the Claude Skills format is built on, which is why skills built for Claude also run in tools like Cursor and Copilot.
Things they pointed at.
02:48toolSemgrep
03:25toolPyright (Python LSP)
Lines you could clip.
00:00
“The number one most downloaded Claude skill on the planet has been installed 829,000 times.”
02:23
“You have a one in eight chance of downloading a potentially malicious skill.”
12:07
“A prompt asks Claude to think. A skill actually teaches Claude how to work.”
24:26
“Every skill you add eats context window and piles on more slash commands. The sweet spot experts agree on is three to five skills.”
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.
metaphoranalogystory
The number one most downloaded Claude skill on the planet has been installed 829,000 times. That same skill was just sitting at 250,000 installs a couple months ago.
So now while you've been using Claude like a fancy chatbot, an entire army of vibe coders quietly turned it into an automated work system. The problem is, there are now over 85,000 community built Claude skills floating around the internet, but 99% of them are actually hot garbage.
So I spent days digging through the entire official Anthropic directory to find the latest install data, and after reviewing over 10,000,000 data points, what I found actually shocked me. 12 skills were consistently showing up more than the rest.
This left me with two options. We can either adopt these, or watch everyone around us work twice as fast for the rest of the year. I know what I'm choosing.
So let me break down all 12 skills for you, starting from number 12 and working our way to the most installed skill on earth. Then, I'll end on the one rule that stops you from nuking your own setup when using too many of these. And no, you do not need to know how to code for any of this.
We're doing everything in natural language. By the end, you'll know exactly which skills to install so you don't get left behind. Oh, and if we haven't met yet, I'm Doobie, former marketer turned vibe coder.
I made over $40,000 in less than fifty days using these exact AI tools, so trust me, the skills I'm about to share, I've been using for weeks now. Starting with number 12, security guidance.
A 175,000 installs. Did you know that according to a recent security audit, roughly 11,000 out of the 85,000 skills available are believed to contain security issues, that means you have a one in eight chance of downloading a potentially malicious skill.
That's why it's more important than ever to keep your security tight. That's exactly what the security guidance skill was built for. It only launched a few weeks ago, and has since skyrocketed straight into the top 12 most installed skills ever.
Here's what it does. It checks every file edit, every commit, every move Claude makes for security holes before they even happen. It actually maps the problems to real industry frameworks like OWASP, the same framework used by billion dollar software companies.
Using these frameworks, it flags risky operations before they ever run. The more autonomous Claude gets, the easier it is for it to quietly introduce a vulnerability you'd never catch. Because of this, enterprises are now treating this skill as mandatory, not optional.
And pro tip, there's this company called Semgrep that specializes in cybersecurity, who have a range of free skills around security. If you pair the security guidance skill with the Semgreps set of skills, you can run much deeper reviews that would easily put your code security in the top 1% of Vibe coders.
If you thought that was cool, wait till you see number 11, the TypeScript LSP a 177,000 installs.
LSP sounds scary. Trust me, it's not. It just gives a real time map of your entire code base.
Every type, every reference, every import. Think of it as handing Claude a GPS for your project, instead of making it read every file blind. And this is the big deal, because it dramatically cuts down on Claude's AI hallucinations.
And honestly, experts say you should install the LSB skill for your main language before anything else. TypeScript tops the list, but there's actually a whole family of coding languages. You've got Python's Pyrite, which sits at 91,000 installs, then go Rust, Java, all the way down.
Whatever you build in, there's one for you. Just ask Claude what's best for your task. 10, Featured Dev.
217,000 installs. This is a full agentic workflow for building a new feature end to end.
It runs a clean five step pipeline, explores the code base, designs the architecture, writes the code, generates the test, then reviews everything. So instead of one claw juggling everything, you've got a little team working in sequence.
I actually pointed this at a feature for one of my apps the other day, and I walked off to make a coffee. A few moments later, I came back to a finished tested feature sitting there waiting for me. For a solo vibe coder, .
Number nine, claude.mdmanagement. 223,000 installs.
Now this one solves Claude's most annoying flaw. Every new session Claude forgets your stack, your naming rules, your conventions, all of it.
The Claude dot m d file is basically Claude's long term memory for a project. This skill helps keep that memory file updated automatically as your project grows. So you stop re explaining the same thing every single morning.
Just set it up once, and Claude remembers who you are. Alright. Between you and me, if you thought that was useful, let me tell you about Number eight, Playwright.
248,000 installs. This gives Claude the power to autonomously control its own headless browser, meaning it can literally search the web, or even better, test out the features of your app.
It literally spins up a browser, clicks through your UI, checks everything works, and fixes the tests when they break. I'm not gonna lie, writing tests is the part of building I actually hate the most. So handing it off to Claude off the rip honestly felt a little wrong.
And I'm not the only one that hates doing it. The creator of Claude, Boris Churney himself, literally created a custom skill called slash go, so that he can quickly tell Claude to use Playwright to autonomously test his own code, allowing him to focus on the finer things in life. Number seven, the GitHub MCP.
261,000 installs. Quick term, MCP just means model context protocol.
In plain English, it's a bridge that connects Claude to a live outside service. This one plugs Claude directly into Github, allowing you to use its features without you ever leaving your workspace. This is actually the most installed connector of all of them, way ahead of Vercel, Figma, and Slack.
It turns Claude from a local helper into an actual member of your team that can actually make changes on your project repo. Fancy that. Number six, Skill Creator, 283,000 installs.
Believe it or not, this one's actually my favorite, because it's the skill that builds skills. What did he say? You can imagine how powerful this thing is.
You describe a task in plain English, and Skill Creator drafts, tests, and packages it into a brand new reusable skill. And it's officially backed by Anthropic. People were saying it's the first skill that you should actually ever install.
Why? Because you're no longer stuck with whatever's pre built. Here's what I recommend.
After a good session, just tell Claude, based on this conversation, build me a skill so we can do this faster next time. Boom. Free custom skill.
And if you're already finding this video useful, do me a favor and smash that like button. It genuinely helps me out. I'll have all the links to every skill in the description below.
Number five, code simplifier. 284,000 installs.
Here's an uncomfortable truth. AI generated code, including Claude's own, is often bloated and over engineered. This skill runs a compression pass on it.
It strips out the redundancy, simplifies the logic, and cuts technical debt, all without changing what the code actually does. Think of it as Claude cleaning up after itself.
For vibe coders shipping fast, this is how you stop your app from turning into spaghetti. Four, code review. 347,000 installs.
This automates the entire code review process. Claude analyzes your changes, checks them against your standards and the original task, then spits out a structured review report. It catches issues before a human ever has to look.
It writes your pull request descriptions, it flags architecture problems early. A lot of people stack it with the PR review toolkit, which adds a confidence score to every single finding.
Genuinely one of the biggest time savers on this entire list. Number three, Context seven. 348,000 installs.
Remember how I said Claude forgets? Context seven is the heavier duty fix. It actually gives Claude live documentation lookup and smart search across huge code bases.
So instead of only knowing what fits in its own context window, it can pull the exact docs and patterns it needs on demand. In big projects, this is the difference between Claude following your conventions, and Claude inventing complete nonsense. This is actually one of three skills experts say you should install globally on every project.
Hold that thought because the other two are coming up right now. The second most downloaded skill, superpowers, 752,000 installs.
This is not a single skill, it's actually an entire framework. It forces Claude to follow a disciplined development methodology, like a senior engineering team.
Here's the problem it fixes. AI agents are eager. They jump straight into writing code before they even understand what you actually want.
Superpowers actually stops this from happening. It makes Claude brainstorm first, break the work into tiny atomic tasks, and follow strict test driven development. If Claude writes code before writing a failing test, the framework forces a restart.
It was built by a developer named Jesse Vincent, and the repo is one of the most AI coding frameworks in the world. Even experts like Simon Willison point to it as the fix for AI's hallucinated enthusiasm problem. Hey, I'm curious.
Are you already paying for Claude or Codex? I recently dropped Ace, an app that helps you get more out of your AI subscriptions by making them work together as a team.
Like, check this out. One AI can research, another can build, and another can check the work. You can make simple teams like this, where Claude and Codex work together to find the best answer, or you can build something much bigger like this autonomous app factory that keeps building apps while I sleep.
Inside Ace, can give tasks to your team in one group chat, Create loops that run tasks again and again, or let one powerful AI manage a group of cheaper ones so you can spend less on tokens. You no longer need to code all of that from scratch. Just drag, drop, and connect the dots.
Now live on Mac and Windows, use code doobie for 10% off. I'll leave a link in the description.
Back to the video. Alright. Now before I reveal number one, I need to give a few honorable mentions, because not every useful skill made in the official top 12, but some of these are genuinely insane.
First up, caveman. And yes, it does exactly what it sounds like. It makes Claude talk like a caveman.
But weirdly, this is actually useful, because Claude loves to yap. It explains every tiny step like it's writing a thesis. Caveman strips all that away.
Less filler, less waffle, same output. So instead of Claude saying, I'll now inspect the relevant files, it just says, me check files, like a caveman. The creator claims it can cut output tokens by around 75%, which means faster sessions, cheaper usage, and way less mental clutter.
It sounds like a meme, but it solves a real problem. Next, G Stack. This one was made by Gary Tan, the CEO of Y Combinator, and it's basically his full Claude code setup in one repo.
CEO review, engineering review, design review, quality assurance, security, release management, docs, everything. Instead of letting Claude randomly build and pray, g stack turns it into a proper startup workflow. Think, plan, build, review, test, ship, reflect, and that matters.
Because once you start building real apps with AI, the hard part is not getting code written, the hard part is actually judgment. Should we build this app? Is the scope too big?
Does the design suck? Did we actually test it? G Stack helps Claude think through those layers before it cooks your project.
Next up, Kapathi skills. Andre Kapathi, the co founder of ChatGPT and head of AI at Tesla, once wrote a thread complaining about how Claude writes code. He mentioned three failure modes.
Forrest Chang read the thread and packaged the complaints into four behavioral rules inside a single Claude dot m d file, and dropped it on GitHub. Currently sitting at a 160,000 stars, it became the fastest growing single file repo of 2026.
The file teaches Claude four simple rules. Think before coding, keep it simple, make surgical changes, and define success before touching anything. Which sounds obvious, but this actually fixes one of Claude's biggest problems.
It's way too confident. It assumes things, it over engineers, it changes files you never asked it to touch, Kapathi skills slows it down, and it forces work with more discipline. So those are the honorable mentions.
Caveman saves tokens, G Stack gives Claude a startup operating system, and Kapathi skills stops Claude from acting like an overconfident intern. Now with that out of the way, let's talk about number one.
The most installed Claude skill on the entire planet with 829,000 installs, it's time to talk about front end design. So why is this king?
Two words, AI slop. You know the look, purple gradients, the same boring font everywhere, generic card layouts, that glassy effect slapped on everything. The glassy effect does look kinda cool.
Every AI generated app looks identical, and developers are sick of it. Front end design forces Claude to commit to a bold, distinctive visual style before it writes a single line of UI. Brutalist, minimalist, retro, futuristic, whatever you want.
And then it enforces it the whole way through. I use this across all of my apps, and the results speak for themselves. Like watch how Tariq, the lead developer at Anthropic, one shot a retro music player.
Genuinely impressive. This is the third member of that infrastructure trio I teased, Frontend Design, Context seven, and Superpowers.
Install those three globally, and Claude gets better on literally every project you touch. Now before you go around crazily installing everything, I have to educate you.
These skills are not prompts. The best builders treat them as standard operating procedures for an employee that never sleeps. A prompt asks Claude to think.
A skill actually teaches Claude how to work. That's the difference. And it actually gets bigger.
These skills run on an open standard called agentskill.io, which means they're not actually locked to Claude at all. The same skill works in Cursor, in Copilot, across basically every AI agent on the market.
You're not learning a Claude trick, you're building a portable workforce that follows you everywhere. But there is a dark side, and you do need to hear it. Anthropic flat out warns that plug ins are highly trusted components that run code with your permissions.
Security researchers have already documented malicious skills trying to silently hijack tool routing in the background. So the rule of thumb is simple. Only install from the official marketplace and treat every skill like a dependency you'd actually vet.
That brings me to the one rule that I promised at the start. Do not install all 12. I know, I know, you want the whole arsenal, but every skill you add eats context window and piles on more slash commands.
The sweet spot experts agree on is three to five skills matched to your workflow. Start with LSP for your language, add the infrastructure trio that I talked about, then pick one or two more for whatever you actually do. Check out this video, YouTube thinks you'll love it.
I'll see you over there.
The Hook
The bait, then the rug-pull.
The creator opens with the punchline number first — 829,000 installs — then works backward to the scale of the problem: 85,000 community Claude Skills exist, and by his account 99% of them aren't worth the context window they eat.
Named ideas worth stealing.
08:03list
The infrastructure trio
- Context7
- Superpowers
- Frontend Design
Three skills the creator says should be installed globally on every project, regardless of what you're building — live docs lookup, disciplined TDD workflow, and enforced visual identity.
Steal forA baseline skill stack to install once on a new machine or project template, before adding anything task-specific.
03:36list
Feature Dev five-step pipeline
- Explore the codebase
- Design the architecture
- Write the code
- Generate the tests
- Review everything
The sequence Feature Dev runs autonomously to take a feature from description to tested implementation.
Steal forA checklist for structuring any multi-step agent task, skill or not.
11:44list
Karpathy's four rules
- Think before coding
- Keep it simple
- Make surgical changes
- Define success before touching anything
Four behavioral rules distilled from Andrej Karpathy's complaints about how AI agents code, packaged into a single CLAUDE.md file.
Steal forA short CLAUDE.md addendum for reining in an overconfident agent.
How they asked for the click.
VERBAL ASK
09:12product
“Now live on Mac and Windows, use code doobie for 10% off.”
Mid-roll sponsor break for Ace (an app for orchestrating multiple AI subscriptions as a coordinated team), woven in as a natural extension of the video's 'stack multiple AI tools' theme rather than a hard interruption; also a soft subscribe/like ask right before it ('smash that like button').
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
Visual structure at a glance.
