Imagine you are setting up a new AI agent to help with your business. Instead of spending three hours writing a massive, fragile system prompt trying to explain exactly how to audit a security vulnerability or how to format a specific type of financial report, you just go to a search bar. You type in "security triage," hit "install," and suddenly your agent just knows how to do it. It is like the App Store, but for actual cognitive abilities.
Herman Poppleberry here, and that is not even a metaphor anymore. We are seeing the birth of a genuine package manager for intelligence. In the last six months, this has exploded from a niche developer experiment into a full-blown ecosystem. We are talking about hundreds of thousands of modular, portable skills that any agent can just "inhale" to become an expert in a specific domain.
It is a wild shift. And speaking of shifting into high gear, today's episode is powered by Google Gemini three Flash. We have a great prompt from Daniel today that really digs into this. He sent over a note saying: Agent skills and skills marketplaces are becoming a big thing in the AI agent ecosystem. There are now dedicated marketplaces where you can browse, share, and install skills for different AI agent platforms. Let's explore what agent skills are, how these marketplaces work, and what it means for the future of AI agents. He specifically pointed us toward Agent Skills dot io, Skills MP, and LobeHub's skills directory.
Daniel is right on the money. We are moving past the era where an AI is just a "chatbot" that you talk to. We are entering the era of the "Agentic Operating System," and these marketplaces are providing the software that runs on top of it.
So, let's start with the basics for a second, because I think people hear "skills" and they think of, I don't know, an Alexa Skill where you ask it to play rain sounds. This is much deeper than that. What is an "agent skill" in the twenty twenty-six sense of the word?
The technical definition has actually stabilized around a standard called "SKILL dot MD." It was originally pushed by Anthropic, but now everyone from OpenAI to Microsoft is using it. At its core, a skill is a bundle. It usually contains a markdown file—that is the SKILL dot MD—which has the core instructions and the "triggers." These triggers are crucial because, unlike an app you have to click, a skill is model-invoked. The AI is monitoring the context of the work, and when it sees a task that matches the "trigger" in the skill metadata, it activates that specific procedural knowledge.
So it is less like me opening an app and more like my brain suddenly remembering a specific training manual the moment I see a specific problem?
That is a great way to put it. And it is not just text instructions. These bundles often include executable scripts—Python, JavaScript, or Bash—that the agent can run to actually perform the action. If the skill is "Optimize React Component," it doesn't just tell you how to do it; it has the linting rules and the performance profiling scripts ready to go.
And we are seeing these pop up everywhere. Daniel mentioned three big ones. LobeHub seems to be the giant in the room right now. I was looking at their stats this morning, and they have over two hundred fifty-four thousand skills listed. That is an insane amount of collective "knowledge" being packaged up.
It is staggering. LobeHub has really leaned into the community aspect. They have categories for everything from "Agent-to-Agent Protocols" to very specific DevOps tools. But then you have Skills MP, which is even larger in terms of raw indexing. They are sitting at over seven hundred forty thousand skills. What they are doing is essentially "indexing the world's procedural knowledge" by scraping GitHub for anything that follows that SKILL dot MD format.
Seven hundred forty thousand. I mean, Herman, how many things are there even to do in the world? At some point, aren't we just getting forty thousand different versions of "how to write a polite email"?
You'd think so, but the granularity is what makes it powerful. On Skills MP, they have "Occupation-based filtering." So you aren't just looking for a "coding skill." You are looking for a skill specifically designed for "Computer and Mathematical Occupations" or "Healthcare Support." It is about the "long tail" of expertise. There might be a skill specifically for "Auditing Smart Contracts for Reentrancy Vulnerabilities on the Solana Network." That is not something you want a general-purpose LLM guessing at. You want the specific, verified procedure.
I see. So it is moving from "General Intelligence" to "Specific Competence." I like that. It feels more reliable. But I'm curious about the execution. If I'm using something like Agent Skills dot io, which seems to be the more "official" documentation and spec hub, how does a platform like Claude Code or OpenAI Codex actually "ingest" these? Is it literally just a download?
In many cases, yes. It functions exactly like a package manager. If you are using a tool like Claude Code, you might run a command like "install skill search-web." The tool fetches the SKILL dot MD and any associated tools from the marketplace, verifies the hash for security, and adds it to the agent's "active library." From that point on, whenever you ask the agent to find something on the web, it isn't just "searching"; it is using a structured skill that might have parameters for search depth, result filtering, and multi-source synthesis.
Okay, so let's talk about that "search web" example because that sounds pretty basic. Why do I need a "skill" for that? Can't the AI just... search?
It can, but think about the difference between a high schooler googling something and a professional researcher. A "Search Web" skill from a place like Agent Skills dot io might integrate with specialized tools like Firecrawl or Exa. It might have a multi-step procedure: first, find five diverse sources; second, extract the raw markdown; third, cross-reference the facts; and fourth, flag any contradictions. By packaging that as a "skill," you ensure the agent follows that high-quality process every single time, rather than just taking the first result it finds.
Right, so it is about consistency and "best practices" being baked into the tool. Now, one thing that caught my eye in Daniel's notes was the "Security Paradox." This is where my "sloth-sense" starts tingling. If we are letting agents autonomously "install" skills to solve problems, aren't we just opening the door for some incredibly creative malware? I can imagine a "helpful" skill that says it optimizes your code but actually just exfiltrates your environment variables to a server in a basement somewhere.
You are hitting on the biggest bottleneck in the entire industry right now. Security. If an agent can self-improve by downloading new capabilities, it can also self-destruct by downloading a Trojan horse. This has led to a really fascinating "meta-market" within these platforms. We are seeing the rise of "Vetter Skills."
Wait, so we have agents whose only job is to audit the skills of other agents?
Precisely. Well, not "precisely," I shouldn't say that... but that is exactly what is happening. There are skills like "skill-vetter" or "prompt-guard" listed on LobeHub. When you try to install a new skill, your "Security Agent" runs the code through a vetter skill first. It reads the instructions, analyzes the scripts for any obfuscated network calls, and looks for "prompt injection" vulnerabilities where the skill might try to hijack the agent's core instructions.
It is "agents all the way down," Herman. It is like a digital version of the "Who watches the watchmen?" dilemma. I find it funny that the solution to "AI might do something bad" is "let's build another AI to watch it."
It is the only way to scale! No human can manually audit seven hundred thousand skills. But what's really cool is the "Agent-to-Agent" or A2A protocols. This is the next frontier Daniel mentioned. Some skills aren't just for doing a task; they are for "hiring" or "delegating" to another agent.
Explain that to me. Give me a real-world scenario.
Okay, imagine you have a "Project Manager Agent." It has a skill for "Software Development Lifecycle Management." You tell it to build a new feature. Instead of trying to write all the code itself, the PM agent uses an A2A skill to "spawn" a specialized "Coding Agent" that has the "Next dot JS Optimization" skill installed. The PM agent delegates the coding, waits for the result, and then perhaps spawns a "QA Agent" with a "Security Audit" skill to check the work. You are essentially building a recursive, automated workforce.
That sounds like a dream for a sloth, Herman. I could just be the "Grand Overseer Agent" at the top of the pyramid, occasionally nodding while my legion of specialized agents does all the heavy lifting. But seriously, this feels like it changes the "moat" for companies. It used to be "we have the best AI model." Now it feels like "we have the best library of curated skills."
That is a massive shift. We've talked before about how the "vibe coding" era is ending. It's moving toward "procedural playbooks." If you are a company like Vercel or Snowflake, your value isn't just in the compute; it's in the specialized skills you provide for your specific ecosystem. Vercel recently released a package manager for AI coding agents specifically tuned for their infrastructure. They want their agents to be the best at deploying to Vercel, so they provide the "Vercel-Expertise" skill as a modular unit.
It makes me think about the job market, too. I saw a stat in Daniel's research about a fifty-six percent wage premium for workers who can manage these agentic systems. It's not about knowing how to code anymore; it's about knowing how to "curate the library."
It is the shift from "Individual Contributor" to "System Architect." If you can go into a marketplace like Skills MP, find the exact five skills needed to automate a complex marketing workflow, and string them together safely, you are worth way more than someone who is just "good at prompting." You are building an engine, not just talking to a machine.
Let's look at some of the more "exotic" skills for a second. Beyond coding and search. What are people actually building?
There is some really specialized stuff. I saw a skill for "Managing WeChat Official Accounts," which involves very specific API calls and content formatting rules that a general LLM would almost certainly hallucinate. There are skills for "Generating PowerPoint Decks" that don't just write the text, but actually use a CLI to build the XML structure of the slides. There is even a category for "Self-Improvement" skills. These are fascinating—they allow an agent to analyze its own previous errors, write a "correction procedure," and save it as a new, local skill so it doesn't make the same mistake twice.
That is slightly terrifying, Herman. An agent that can literally learn from its mistakes and write its own "upgrade" code. We are one step away from them realizing that humans are the ultimate "inefficiency" and writing a skill for "Human Management."
Well, let's hope the "Vetter Skills" are strong enough to catch that one! But honestly, the practical side is where the value is. Think about "Deep Research." Most people use AI to summarize an article. But a "Deep Research" skill might integrate with things like Firecrawl or Exa. It can perform multi-source synthesis across dozens of PDFs, academic papers, and live web data. It turns the agent into a world-class analyst for the cost of a few API credits.
And the "Package Manager" model means this is all composable. I can take a "Deep Research" skill from one developer, a "Financial Reporting" skill from another, and a "Strategic Analysis" skill from a third, and suddenly I have a "Chief Financial Officer" agent.
And you can swap them out! If a better "Financial Reporting" skill comes along, you just update the package. You aren't tied to one monolithic prompt. This modularity is what is going to drive the "$100 Billion AI Agent Economy" that the analysts are predicting for the next few years.
So, if I'm a listener and I'm hearing this, and I'm thinking, "Okay, this sounds cool, but where do I actually start?" What is the "hello world" of agent skills?
I would say go to Agent Skills dot io first. It is the cleanest jumping-off point. They have a list of all the major industry adopters. If you are using Cursor, or Claude Code, or even something like Snowflake Cortex, you are already in the ecosystem. Look at the "Official Adoption" list. Then, head over to LobeHub. Their interface is very user-friendly—it feels like a modern app store. Browse the "Coding" or "General" categories. Just look at the SKILL dot MD files. They are surprisingly readable. You'll see exactly how the "triggers" work and how the instructions are structured.
It’s basically just a very well-organized set of rules, right? It isn't magic. It's just... extreme organization.
It is "Procedural Knowledge" made portable. That is the key. Before, that knowledge was trapped in a human's head or a very long, messy text file. Now, it is a standardized unit of software.
I want to go back to the "Second-Order Effects" you mentioned earlier. If every developer can now monetize their specialized expertise by selling "skills" on these marketplaces, do we see a "Skill Economy" emerging? Like, could I make a living just writing the world's best "Legal Compliance for European Privacy Laws" skill?
I think that is a certainty. We already have the "Creator Economy" for influencers. The "Skill Economy" is for domain experts. If you are a top-tier lawyer, or a specialized engineer, or a niche researcher, your "product" is no longer your time—it's your "Procedure." You can package your unique way of solving a problem into a skill, put it on a marketplace, and get a micro-payment every time an agent uses it to solve a problem for someone else.
That is a much better way to scale expertise than "consulting." But there’s a flip side. If I can just "buy" the skill of a top-tier lawyer for my agent, what happens to the junior lawyer who used to do that work?
That is the big question. The "entry-level" procedural work is the first thing to be packaged into these skills. If an agent can do "Initial Document Review" using a high-quality skill from a marketplace, the junior associate who used to do that is in trouble. This is why that "fifty-six percent wage premium" for managing these systems is so important. You have to move up the value chain. You aren't the "doer" anymore; you are the "curator and auditor."
It feels like we are moving toward a world where the "Human-in-the-Loop" is more like a "Human-on-the-Loop." We are standing outside the machine, swapping out parts and making sure it doesn't overheat.
And that brings us to the "Agent-to-Agent" stuff again. One of the most interesting things I saw on Skills MP was a category for "Negotiation Protocols." Skills that allow two different agents, representing two different people, to negotiate a schedule or even a contract price using a standardized set of rules.
Oh, I would love that. I hate negotiating. If I could just have my "Sloth Agent" talk to your "Donkey Agent" to decide who's paying for dinner, that would save me a lot of social anxiety.
But think about the complexity there! The agents need a shared "language" for that negotiation. That is what these skills provide. They aren't just "instructions"; they are "protocols." They are the "TCP-IP" of the agent world.
Okay, let's get practical for a minute. We've talked about the big picture. If someone is building an agent today, say for a small marketing team. They want to automate their social media presence. Instead of writing one giant script, how should they approach it using the "Skills" mindset?
Step one: Break the job down into "Atomic Procedures." Don't try to build a "Social Media Agent." Build an agent that uses five skills. One skill for "Trend Analysis" using a web search tool. One skill for "Drafting engaging X posts" based on a specific brand voice. One skill for "Image Generation" using a DALL-E or Midjourney API. One skill for "Scheduling and Posting" via a CLI. And crucially, one skill for "Performance Review" that looks at the analytics and feeds the data back into the first skill.
And the beauty is, if the "Image Generation" skill gets better next month, you just swap that one module. You don't have to rewrite your entire "Social Media Agent."
And you can find most of those already pre-built. You might go to LobeHub and find a "Twitter CLI" skill that has been tested and "vetted" by ten thousand other users. That is much safer and more reliable than trying to write your own integration from scratch and hoping you didn't leave a security hole.
I like the "community-vetted" aspect. It's like open-source software, but for "ways of thinking." But I do wonder about the fragmentation. If I buy a skill on LobeHub, will it work on my agent that I built using a different framework?
That is where the "SKILL dot MD" standard is so vital. Because it is essentially just Markdown and standardized JSON for the tools, it is remarkably portable. Skills MP is a great example of this—they are framework-agnostic. They index skills that can work across four or five different agent platforms. We are seeing a move toward "Interoperability." It is in no one's interest to have "walled gardens" for intelligence.
Well, maybe not in the users' interest, but I’m sure some big tech companies would love a walled garden. "Only OpenAI-certified skills work on our agents."
They might try, but the open standards are moving too fast. When you have seven hundred thousand skills already in the wild using an open format, trying to lock it down now is like trying to lock down the internet after the invention of HTML. The cat is out of the bag.
So, we have these massive repositories. We have the security vetters. We have the agent-to-agent protocols. Where does this go in the next twelve to eighteen months?
I think we start seeing "Skill Bundles" for specific professions. You won't just buy a "skill"; you'll subscribe to the "Certified Public Accountant Suite" for your agent. It will be a curated, constantly updated collection of dozens of skills that keep your agent up-to-date with the latest tax laws and reporting standards.
And the "human" CPA is the one who "signs off" on the final output, but the agent does ninety-nine percent of the procedural work using those skills.
It’s the "Cyborg Professional." The combination of human judgment and agentic precision. And it all goes back to what Daniel was saying—the marketplace is the infrastructure that makes this possible. It turns "AI development" into "AI assembly."
I'm looking at Skills MP right now, and the "Occupation-based filtering" is actually really deep. They have sections for "Architecture and Engineering," "Life, Physical, and Social Science," even "Legal." It’s like a map of human labor, but digitized.
It is. And what's interesting is which categories are growing the fastest. Right now, "Development and Coding" is still the king, but "Security and Vetting" is the fastest-growing niche. People are realizing that as we give agents more power—the ability to run code, to manage cloud infrastructure, to move money—the "Skills" they use need to be bulletproof.
I saw a skill on there called "skill-vetter-pro." I love the idea of a "Pro" version of a tool that watches other tools. It’s like the "Premium Security Guard" for your digital workforce.
It’s necessary! If you are a big enterprise, you aren't going to let your agents download random scripts from GitHub. You are going to have a "Private Skill Marketplace" within your company. You'll take the best skills from the public marketplaces, have your security team audit them, and then put them in your internal "Approved Library."
"Shadow IT" becomes "Shadow Agents." "Oh, I just downloaded this cool skill that summarizes my meetings!" "Wait, did you just give our entire product roadmap to a third-party script?"
That is going to be the biggest headache for IT departments in twenty twenty-six. Managing the "Skill Permissions" of their employees' agents. It’s a whole new layer of governance.
Okay, let's wrap this part of the discussion up with some takeaways. I think we've established that the "Skill" is the new "App." If you're a developer, you should be looking at the SKILL dot MD standard. If you're a user, you should be exploring these marketplaces to see what you can "outsource" to your agents.
My main takeaway is: Stop thinking about AI as a "box you talk to." Start thinking about it as a "frame you build on." The marketplaces like Agent Skills dot io and LobeHub are giving you the bricks to build something much more powerful and reliable than just a clever prompt.
And don't forget to have a "Vetter Agent" watching your "Worker Agent." Trust, but verify... and then have an AI verify the verification.
That should be the slogan for the twenty-first century.
So, let's get into the "Practical Takeaways" for our listeners. Because this stuff can feel a bit "future-y" and abstract, but it's happening right now. Herman, if someone wants to actually do something with this information today, what's step one?
Step one is a "Skill Audit" of your own workflow. Look at the tasks you do every day that are "procedural." Things like "Reviewing a pull request for style guide compliance," or "Formatting a weekly status report," or "Searching for competitors' new feature announcements." These are perfect candidates for skills.
And then you go to the marketplaces.
Right. Go to Skills MP dot com or LobeHub. Search for those specific tasks. I guarantee you'll find at least three or four skills that already do seventy percent of what you need. Download them, look at the SKILL dot MD file to see how they work, and try "installing" them into your agent framework of choice. Even if you're just using a tool like Cursor or Claude Code, you can often "feed" it a skill file as part of your project instructions.
My takeaway is for the business owners and managers. If you are hiring people right now, or training your team, you need to be looking for "Agent Orchestrators." The person who knows how to curate a high-quality library of skills is going to be ten times more productive than the person who is just "good at their job" in the traditional sense.
It is about "Leverage." A skill is a piece of leverage that doesn't sleep and doesn't forget.
And finally, don’t ignore the security side. If you are going to use these marketplaces—and you should—make sure you understand the "Vetter" model. Don't just install a skill from a random developer without at least having another LLM read the code and explain what it's doing. It takes thirty seconds and could save you a massive data breach.
"Model-on-Model auditing." It’s the only way to stay safe in a world of seven hundred thousand autonomous skills.
I think that’s a great place to leave it. This has been a fascinating look into the "Package Manager for Intelligence." It feels like we are finally moving away from the "magic trick" phase of AI and into the "industrial tool" phase.
The "Great Assembly" has begun.
Before we wrap up, I want to say thanks as always to our producer, Hilbert Flumingtop, for keeping the gears turning behind the scenes. And a huge thank you to Modal for providing the GPU credits that power this show—we literally couldn't do this without that serverless horsepower.
This has been My Weird Prompts. If you're finding these deep dives helpful, we'd love it if you could leave us a review on Apple Podcasts or Spotify. It’s the best way to help other "curious humans" find the show.
We'll be back next time with another prompt from Daniel. Until then, stay curious, and maybe go find a skill that handles your emails so you can take a nap.
Or a skill that teaches you how to be a better sloth?
I already have the "Master" version of that one, Herman. It’s built-in. See ya!
Goodbye!
