Insane Claude Skills That Actually 10x Your Vibe Coding

AAI LABS
Computing/SoftwareAdvertising/MarketingSmall Business/StartupsInternet Technology

Transcript

00:00:00Agent skills make your AI agents so much more powerful. This is why we're bringing you six
00:00:05amazing skills that help you build much faster with AI agents. But these skills aren't going to
00:00:10replace the way you already work. If you already have a way of working with your agent, you can
00:00:14keep using it. These skills won't change that. Instead, they'll add new things your AI coding
00:00:19agent can do that you may not have thought were possible before. These skills were made by some
00:00:24of the most well-known people in AI. For example, one of them gives us four rules that we use in
00:00:29every single AI coding project. Another completely changes how AI agents see the internet and use
00:00:35websites. Now, if you're new to the channel, then welcome. We're a software company and this is AI
00:00:40Labs. And in this video, we're going to go over six skills that will definitely change how you work
00:00:45with AI agents. Now, the biggest problem with skills isn't making them. It's keeping them updated. Every
00:00:50time your agent repeats a mistake, you have to notice it and improve the skill yourself. But the task
00:00:55observer skill does this while you're actually working. The whole idea behind this skill is that
00:01:00skills get better from the problems you run into during real work. You can't sit down once and
00:01:05predict every mistake the agent is going to make. So it watches for useful lessons while you work. If
00:01:10the same solution could become a new skill, it records that. And if an existing skill needs a new rule,
00:01:15it records that as well. When you install this skill, you have two options. You can install it in
00:01:19the main.clawd folder, which makes it work across every clawd code session on your computer. Or you
00:01:25can install it inside one project. We recommend installing it per project because that keeps the
00:01:29lessons for each project separate and makes them easier to manage. Once it's installed, it creates
00:01:34a file called log.md, where everything it notices gets recorded. This file sits inside clawd code's main
00:01:40folder on your computer outside the actual project. If you install the skill per project, each project
00:01:45keeps its own log. Next to that is another file for rules that should apply to every skill you own.
00:01:51So let's say you decide that none of your skills should ever write a real client's name into a file.
00:01:56That rule isn't only for the skill you were using at the time, it should apply to all of them. So it goes
00:02:00into this second file. During the setup, this skill also asks you to add an instruction to your clawd.md
00:02:06file. That instruction makes clawd run the skill while you're working so it can keep finding lessons and
00:02:11improvements. Inside the log, every skill gets its own section and the suggested changes are written
00:02:16underneath it. But this skill doesn't immediately change the real skill. The lessons in the log are
00:02:21used during future sessions and you can review them before deciding whether they should become a
00:02:26permanent part of the skill. This stops one bad result from automatically becoming a rule that the
00:02:31agent follows forever. We ran into this with the skill we use to create animations for our videos.
00:02:36That project has a lot of custom instructions that make the animations look the way they do and we're
00:02:42constantly improving them. A lot of those animations include brand logos. Models can't create those
00:02:47logos properly on their own. So we need to give them the logo in a code format called SVG. We had been
00:02:52getting those files manually from a site called LoebHub. LoebHub also has a tool that agents can run from
00:02:58the terminal so our agent assumed it could use that tool to get the logos but the tool was made for
00:03:02something else so it failed. This skill recorded that the LoebHub tool should not be used for our
00:03:08animation system. It also found the correct source for getting the logos and recorded that as a new
00:03:13rule. We reviewed both lessons and added them to the skill. Now the animation agent knows exactly where
00:03:18to get the brand logos and can add them without us doing it manually. But before we move on to the next
00:03:23skill, it would be great if you subscribe to the channel and hit the hype button. This small gesture of
00:03:28support goes a long way for us. Your agent can build a working product but a working product can still
00:03:34lose people before it makes any money. These next skills cover the parts of your app that stop that
00:03:39from happening. These skills come from Corey Haynes and they're part of a collection of 48 marketing
00:03:44skills. The name might make them sound like they're only for ads and content but the collection covers
00:03:49all the different ways a product makes money and three of those are built directly into the app.
00:03:54First, we have a really important part of the app that most people coding with AI completely miss
00:03:59and that's onboarding. Onboarding is the process that helps a new user set up and start using your
00:04:04product. Then we have the main thing that drives your conversion which is your paywall. A paywall
00:04:09is the screen that appears when a user tries to use something they need to pay for. And third,
00:04:14we have churn which is one of the most important numbers for understanding whether people stay with
00:04:18your product. Churn is when people who are paying for your product stop paying. So this is the app
00:04:23we're going to use for the rest of the video. It's a next.js app which is one of the standard ways to
00:04:28build a web app and we've built it for a gym. People can use it to get a gym membership, pick their
00:04:33trainer and choose their exercises. Basically everything needed to run the gym is inside one
00:04:38app. For the database, we're using Superbase which is one of the easiest ways to store everything the
00:04:43app needs. It stores the users and all the information inside the app and it also handles the login
00:04:48system. If we open the skills folder, you can see the three marketing skills alongside the Superbase
00:04:53skills. And this is what a skill is for. You use it to give your agent information that it doesn't
00:04:58already have. Superbase is constantly being updated so the model's training doesn't always match the
00:05:03version you're using. It's also one of the most popular databases for apps like this. So if you're
00:05:08building with it, these skills are worth installing. We'll link them in the description as well.
00:05:12Starting with the onboarding skill, what you're looking at here is a prototype. That means it
00:05:16hasn't been added to the real app yet. It's only here to show us what the skill came up with. Before
00:05:21we use the skill, signing up meant answering five different things across five separate steps. That
00:05:26made it much harder for someone to get inside the app and start using it. So the new version fills in
00:05:31the person's name from the card they already completed. When they create their account, the only new
00:05:35thing it asks for is the location of their gym. After that, it takes them straight to the plans and
00:05:40already has the most popular one selected, which makes the decision easier. These aren't made up
00:05:45choices. They're based on how people have behaved across real products over years of testing. Once
00:05:50the user enters the app, every setup step we removed is moved into a progress bar along the bottom. The
00:05:55app only asks for the information it needs to get them started and the notification stays there until
00:06:00they finish the rest later. But getting somebody into the app doesn't mean they'll pay for it. That's
00:06:05where the paywall skill comes in. The approach this skill took was to let people sign up for free
00:06:10and then show them what they would get if they paid for a plan. Some of the content is available,
00:06:14while the rest stays locked. The screen then sends them directly to the pricing page where they can
00:06:19unlock it. You've seen this pattern in plenty of apps before, but if you simply ask an agent to add a
00:06:24paywall, it normally won't make these choices because it doesn't have these principles built into it.
00:06:29And getting someone to pay still doesn't mean they'll keep paying. So the last skill is about
00:06:33preventing churn. For this one, we've got the cancellation page. When someone tries to cancel their membership,
00:06:38the app doesn't immediately end it. It first asks why they want to leave. Once they choose a reason,
00:06:44it gives them a better offer based on that answer. The screen also uses smaller design choices to reduce
00:06:49cancellations. Pausing the membership is the main button on the page, while continuing with the
00:06:53cancellation is shown as plain text underneath it. Now this implementation is still generic because
00:06:58it's a demo. Once you give the skill your real user data and the details of what you're selling,
00:07:03its decisions can become much more specific to your product. Most problems with coding agents don't come
00:07:09from the code itself. They come from the agent making decisions you never asked it to make. This next
00:07:14skill turns Andrej Karpathy's advice into four rules that stop that from happening. The first rule stops the
00:07:20agent from guessing when your request isn't clear. It has to tell you what it's assuming, show you the
00:07:25different options, and ask before it makes a decision that could take the project in another direction.
00:07:29The second rule stops it from building more than you asked for. It has to use the simplest solution
00:07:35that works instead of adding extra features, settings, or systems for problems you don't have
00:07:39yet. The third rule stops it from changing anything outside the job you gave it. It has to match what's
00:07:44already there, only clean up the mess it created, and tell you about other problems instead of fixing
00:07:49them on its own. And the last rule stops the agent from calling a job finished just because it changed
00:07:54something. Before it starts, it has to decide what the result should be and how it will check that result.
00:07:59Then it has to keep working until every check passes. Now you can install this as a normal skill,
00:08:04but we use it a little differently because these are rules we want inside every coding project. On our
00:08:09Macs, we keep all our coding projects inside one developer folder. We've given that folder its own
00:08:15custom icon and dragged it into the finder sidebar so every project starts from the same place.
00:08:20Inside that developer folder, we don't install this as a skill. We copy the four rules into a
00:08:25clawed.md file at the root of the folder. Clawed code reads the clawed.md file inside a project and
00:08:31any clawed.md files in the folders above it. So whenever we open a coding project inside the
00:08:37developer folder, these four rules are loaded alongside that project's own instructions. This
00:08:42gives every coding project the same base rules without applying them to unrelated clawed code sessions on
00:08:48the rest of the computer. One of the biggest limits of AI agents is that they're still bad at using
00:08:53websites. Browser agents have to open every page, find the right buttons and slowly work through the
00:08:58site. OpenCLI changes that by turning more than 100 websites, including the accounts you're already
00:09:04logged into, into commands your agent can run directly. Now this one isn't just a skill, it has
00:09:08three parts that work together. You have the main OpenCLI tool, an extension you add to Chrome and a set
00:09:14of skills that teach your agent how to use everything. OpenCLI already knows how to do different jobs
00:09:19across more than 100 websites. Each command does one clear thing, so instead of asking your agent
00:09:24to OpenX, find your bookmarks and copy everything out, it can run one command and get your bookmarks
00:09:30back as a clean list. The Chrome extension connects OpenCLI to the websites you're already logged into.
00:09:35This means your agent can use your account without you giving it your password again. The skills then
00:09:40teach your agent how to find the right command, make a new one when needed and fix one if the website
00:09:45changes. A good example is X. X charges for access to its official API, which is the way other software is
00:09:51allowed to use its data. But if you're already logged into X in Chrome, OpenCLI lets your agent search
00:09:57posts, read your bookmarks, post, reply and follow people through your account. So if you're researching
00:10:02a product idea, your agent can search what people are saying about the problem. It can also take all
00:10:08the posts you've saved and turn them into a clean list it can use for research. This is made for jobs you
00:10:13run through websites using your own account. It isn't meant to replace the official tools inside an app you're
00:10:18selling to other people. AI models tend to repeat the same designs. So instead of accepting the first
00:10:24design they give you, you should ask for several versions and compare them. But inside a real
00:10:29project, keeping track of several versions of each part of a page gets messy very quickly. This is where
00:10:34the very 8 skill comes in. It's from the same author who made the hallmark design skill which has gained
00:10:39a lot of stars and has been featured on our channel before. We mainly recommend this skill for landing pages
00:10:45because it makes trying different ideas much easier. So this is our gym app from before. This skill adds a
00:10:50small menu that lets you make and compare different versions of one part of the app. We've added it to
00:10:55the landing page. So as soon as we sign in and enter the main app, the menu disappears. From this menu,
00:11:00we can switch between all the different versions. Right now, we've only made variations of the hero
00:11:05section. So as we switch between them, the rest of the landing page stays exactly the same. Now,
00:11:10this is a real Next.js app, but if you've been watching the channel, you know that when we're
00:11:14designing an app, we normally start with a simple HTML page. So here we have the same type of landing
00:11:20page built in HTML and the very 8 menu works here as well. We can switch between the different versions
00:11:25and choose the one we like. That choice is sent directly to the agent and then we can ask it to
00:11:30make more versions of another part of the page. This means you don't have to redesign the whole landing
00:11:34page every time. You can try different versions of one button, one section or the header and then mix
00:11:40them together until you find the combination you like. Every skill we've covered so far helps your
00:11:45agent build a better product, but none of that matters if you're building something nobody wants.
00:11:50This final collection of skills helps you make the first decision, which is whether you should build
00:11:54the product at all. These skills were made by Sahil Lavindia, who started Gumroad, one of the most
00:11:59popular places for selling digital products online. They're based on his book called The Minimalist
00:12:04Entrepreneur, which has been turned into 10 skills. Nine of those skills take you through the stages
00:12:09of building a business. The first skill helps you find a group of people to build for and it works
00:12:14backwards from how most people come up with an idea. Normally, you pick an idea first and then go
00:12:19looking for people who might want it. But this skill starts with the people. It makes you look at the
00:12:24groups you're already part of and find the problems people keep complaining about. One of those
00:12:29problems can become your product idea. Then it checks if that group is worth building for. It asks whether
00:12:34you can reach those people without paying for ads and whether they already spend money to solve problems
00:12:39like this. Finding a real problem gives you an idea, but it still doesn't prove that anybody will pay
00:12:44for your solution. So the second skill is about validation, which means checking if people will
00:12:49pay before you build anything. It won't accept a broad answer like businesses when it asks who has the
00:12:54problem. It wants 10 real people you can name and it wants at least three of them to say
00:12:58they would pay you. Once you've proved that people want the product, the other seven skills guide you
00:13:03through what comes next. They cover choosing the smallest version, doing the work by hand first,
00:13:08getting your first customers, setting your price, marketing the product, growing without raising money
00:13:12and deciding how the company should run before you hire anyone. Now everything we build and show you in
00:13:18all our videos can be found in AI Labs Pro, which is our community. So if you found value in what we do
00:13:23and want to support the channel, this is the best way to do it. The link's in the description. That
00:13:28brings us to the end of this video. If you'd like to support the channel and help us keep making videos
00:13:33like this, you can do so by using the super thanks button below. As always, thank you for watching and
00:13:38and I'll see you in the next one.

Key Takeaway

Integrating specialized skills into Claude Code agents—such as automated log observers, marketing modules, and design variation tools—transforms raw coding capability into structured product development.

Highlights

  • The task observer skill automatically records lessons and rules into a log.md file during actual coding work.

  • A collection of 48 marketing skills includes onboarding, paywall design, and churn reduction strategies built directly into web apps.

  • Four specific rules derived from Andrej Karpathy stop AI coding agents from guessing, overbuilding, changing unrelated files, or finishing prematurely.

  • OpenCLI combines a command-line tool, a Chrome extension, and agent skills to turn over 100 authenticated websites into direct terminal commands.

  • The very8 skill adds a toggle menu to Next.js apps and HTML pages to generate and compare multiple design variations for specific landing page sections.

  • Ten business skills based on Sahil Lavingia's book require identifying a specific group with a problem and securing commitments from 3 out of 10 named individuals before building.

Timeline

Automating Skill Improvements with Task Observer

  • The task observer skill watches for useful lessons during real work and records them in a log.md file.
  • Installing the skill per project keeps lessons separate and manageable inside specific directories.
  • The log records suggested changes and skill updates for review before they become permanent rules.

Keeping skills updated during coding tasks is difficult because agents repeat mistakes unless manually corrected. The task observer skill monitors workflows to record necessary adjustments automatically. Instead of modifying permanent rules instantly, it writes suggestions to a log file for later review. For instance, when an animation agent incorrectly attempted to use a LoebHub terminal tool for brand logos, the observer logged the failure and the correct SVG source path.

Marketing and App Retention Skills

  • Corey Haynes' marketing collection integrates onboarding, paywalls, and cancellation flows directly into web applications.
  • Streamlined onboarding prefills user names and highlights the most popular subscription plan to reduce friction.
  • Cancellation pages reduce churn by prompting for a leave reason and featuring membership pausing as the primary option.

Building a working product does not guarantee revenue if users abandon the application early. A collection of 48 marketing skills addresses this by embedding core business logic into the app. For a gym app built with Next.js and Supabase, onboarding is shortened by removing multi-step signup barriers. The paywall introduces a free tier with locked features pointing directly to pricing pages. The cancellation flow captures exit reasons and prioritizes pausing memberships over full cancellations.

Enforcing Agent Boundaries with Karpathy Rules

  • Four explicit rules stop AI agents from guessing unclear requests, overbuilding unrequested features, or modifying external files.
  • Placing the rules in a root clawd.md file applies them automatically to every project inside a designated developer folder.

Coding agents often make autonomous decisions that drift away from user intent. Translating Andrej Karpathy's advice into four core rules restricts agent behavior. The rules mandate that the agent must ask clarifying questions for ambiguous requests, build the simplest working solution, restrict edits strictly to the assigned task, and verify results against predefined checks before finishing.

Integrating Websites into Agent Workflows with OpenCLI

  • OpenCLI turns over 100 authenticated websites into direct terminal commands using a Chrome extension and dedicated skills.
  • Agents use existing browser login sessions to search posts, read bookmarks, and gather research data without official API access.

Browser agents are often slow and inefficient at navigating web pages. OpenCLI solves this limitation by combining a main tool, a Chrome extension, and agent skills. The Chrome extension connects to active browser sessions so agents can execute actions through user accounts. This allows developers to query platforms like X for product research and retrieve saved bookmarks as clean lists without relying on paid APIs.

Rapid Design Variations with Very8

  • The very8 skill adds a toggle menu to Next.js and HTML landing pages for comparing multiple design variations.
  • Isolating variations to specific sections like hero elements prevents the entire page layout from breaking during redesigns.

AI models tend to repeat standard design patterns unless prompted with multiple variations. The very8 skill creates an interactive menu that lets developers generate and switch between different versions of a specific page component. Applied to a gym app landing page, it allows testing multiple hero section variations while keeping the rest of the page layout identical until a final choice is sent back to the agent.

Validating Business Ideas Before Building

  • Sahil Lavingia's ten minimalist entrepreneur skills guide founders from audience identification to business launch.
  • Validation requires naming ten real individuals with a specific problem and securing commitments from at least three of them.

Writing code for an unwanted product wastes development effort. The final collection of skills, based on Sahil Lavingia's book, structures the initial validation phase. Instead of starting with a product idea, the first skill identifies existing groups and their complaints. The validation skill enforces strict criteria by demanding ten named targets and at least three paying commitments before any code is written.

Community Posts

No posts yet. Be the first to write about this video!

Write about this video