顧哥知識系統

STOP Building AI Agents. Do THIS Instead.

閱讀時間約 14 分鐘

STOP Building AI Agents. Do THIS Instead.

Transcript

[00:00] Everyone is building AI agents right
[00:01] now. Every startup, every tech company,
[00:03] every developer on Twitter. [music]
[00:05] Agents, agents, agents. And it’s the
[00:07] hottest word in AI. But what if I told
[00:09] you that Anthropic, the company that
[00:11] actually [music] builds Claude, just
[00:13] said, “Stop building agents.” Not me,
[00:15] them. The people building the AI are
[00:17] telling us that the future isn’t more
[00:19] agents, [music] it’s something they call
[00:20] skills. So, in this video, I’m going to
[00:22] break down exactly what skills [music]
[00:24] are, why they matter, and why I think
[00:26] this changes everything, especially if
[00:28] you’re someone who’s been trying to
[00:29] figure out how to actually make AI
[00:31] useful for real work and not just demos.
[00:34] And then, I’m not just going to explain
[00:35] this in theory, I’m going to show you a
[00:37] real project I built [music] using
[00:39] Claude code skills, which are a set of
[00:40] skills that can take any website, run a
[00:43] full AI search optimization [music]
[00:45] audit on it, and generate a professional
[00:47] report you can charge clients for, all
[00:49] powered by skills. [music] So, let me
[00:51] first set the stage. Why would
[00:53] Anthropic, a company that literally
[00:55] builds AI, tell people to stop building
[00:57] agents? Here’s the thing, AI agents
[00:59] right now are incredibly powerful. They
[01:01] can write [music] code, browse the web,
[01:03] call APIs, reason through complex
[01:05] problems. The raw intelligence is there,
[01:07] no one is disputing that at all. But
[01:09] intelligence is not the same as
[01:11] expertise. Think about it like this,
[01:12] imagine you need someone to do your
[01:14] taxes, right? You’ve got two options.
[01:16] Option A, a genius, someone with a
[01:18] sky-high IQ who can figure out anything
[01:21] from first principles. They’ve never
[01:23] done a tax return before, but they’re
[01:24] brilliant, they’ll work it out. Option
[01:26] B, an experienced tax professional. They
[01:29] have done thousands of returns, they
[01:30] know every rule, every edge case, and
[01:33] every deduction. So, who are you picking
[01:35] in this scenario? You’re picking the tax
[01:36] professional [music] every single time,
[01:38] cuz you don’t want a genius figuring out
[01:40] the tax code from scratch, you want
[01:42] someone who already knows it. And that’s
[01:44] exactly the problem with AI agents
[01:45] today. They’re brilliant generalists,
[01:47] they can figure things out if you give
[01:49] them enough time and the right guidance,
[01:51] but they don’t come preloaded with
[01:53] specific [music] expertise. They don’t
[01:54] know your workflows, they don’t know
[01:56] your industry, they don’t remember what
[01:58] worked last time. So, what does everyone
[02:00] do? They build a separate agent for
[02:02] every use case. [music] A tax agent, a
[02:04] legal agent, a marketing agent. Each one
[02:07] with its own custom tools, its own
[02:09] setup, its own architecture. And it’s
[02:11] exhausting. It does not scale at all.
[02:13] What Anthropic realized is that that
[02:15] agent underneath is actually universal.
[02:16] [music]
[02:17] You don’t need a different agent for
[02:19] every job. What you need is a way to
[02:21] give the same agent domain [music]
[02:23] expertise on demand. That’s what skills
[02:25] are. Okay, so let me explain this as
[02:27] simply as possible. At its core, a skill
[02:30] is a markdown file. Basically, a
[02:32] document of instructions that teaches
[02:34] Claude how to do [music] something
[02:35] specific. Think of it like a playbook
[02:37] you would hand a new employee. Here’s
[02:39] how we do things. Here are the steps.
[02:41] Here’s what to watch [music] out for.
[02:42] That file lives inside a folder, and
[02:44] that folder can also include other
[02:46] things like scripts that Claude can run,
[02:48] reference documents, templates,
[02:50] examples, whatever Claude needs to
[02:52] [music] do expert-level work in that
[02:54] area. But the heart of every skill is
[02:56] that instruction file. So, at its
[02:57] simplest, you can create a skill in 5
[02:59] minutes. You write a markdown file that
[03:01] says, “When someone asks you to write a
[03:03] blog post, follow these steps. Use this
[03:05] tone. Structure it in this way. Always
[03:07] include these sections.” Save it in the
[03:09] right folder. Done. Claude now knows how
[03:12] to write blog [music] posts the way you
[03:14] want them to written. But skills can
[03:15] also be much more sophisticated. [music]
[03:17] You can include Python scripts that
[03:19] Claude can execute. You can have
[03:20] templates it fills in. You can build
[03:22] entire workflows. Now, here’s the part
[03:25] that makes this [music] really clever.
[03:26] Skills are what Anthropic calls
[03:28] progressively disclosed. That’s a fancy
[03:30] term for something simple. Claude
[03:32] doesn’t load every [music] skill into
[03:34] memory all at once. When Claude starts
[03:36] working, it just sees a short
[03:38] description of each skill, like reading
[03:40] the title on a spine of a book. Only
[03:42] when Claude decides it actually needs a
[03:44] particular skill for the task at hand,
[03:46] [music] does it pull it off the shelf
[03:47] and read the full instructions. This
[03:49] means you can give Claude access to
[03:51] hundreds or even thousands of skills and
[03:53] it won’t get [music] overwhelmed. It
[03:54] picks the right ones for the job
[03:56] automatically. Just like how an expert
[03:58] knows which reference manual to grab
[04:00] without having to reread all of them.
[04:02] And this is what makes skills so
[04:04] different from just stuffing a bunch of
[04:05] instructions into a prompt. With a
[04:07] prompt, everything is clamped into
[04:09] context [music] at once and it gets
[04:10] messy fast. With skills, it’s organized,
[04:13] it’s modular, and Claude pulls it only
[04:15] when it needs when it needs it. Now,
[04:17] here’s the part that really matters for
[04:19] non-technical people. Before skills, if
[04:21] you wanted to customize AI behavior, you
[04:23] basically needed to be a developer. You
[04:25] had to build custom [music] tools, write
[04:27] code, set up infrastructure. With
[04:29] skills, anyone can create one. You’re a
[04:31] recruiter, write a skill that teaches
[04:33] Claude your company’s hiring process,
[04:35] your evaluation criteria, your interview
[04:37] questions bank. You’re in finance, write
[04:39] a skill that teaches Claude how to build
[04:41] reports the way your [music] team does
[04:43] them. You’re in legal, teach Claude how
[04:45] to review contracts the way your firm
[04:46] does it. And skills also work alongside
[04:49] something called NCP servers, which are
[04:51] Claude’s connections to external tools
[04:53] and [music] data. Think of it this way,
[04:54] NCP lets Claude reach out and interact
[04:56] with the world, pull data from APIs,
[04:58] connect to your tools. Skills tell
[05:00] Claude what to actually [music] do with
[05:02] all of that. NCP is the hands, skills
[05:05] are the experience. Now, when you put
[05:07] these together, you get something really
[05:09] powerful, an agent that can connect to
[05:11] anything and knows exactly how to handle
[05:13] what it finds.
[05:15] All right, so now that you have a little
[05:16] bit of idea of what skills are, let me
[05:18] now show you a practical project so you
[05:20] know exactly what I’m talking about. So,
[05:22] I’m inside my Claude code here and I’m
[05:24] going to run a GEO analysis on a real
[05:27] website. By the way, if you’re
[05:28] interested in learning more about
[05:30] skills, check out the community link is
[05:31] going to be description. I have an
[05:33] entire skills library that you can
[05:34] download for different business use
[05:36] cases, for research analysis. And then
[05:38] also if you’re interested in learning
[05:40] how to monetize this and sell this to
[05:42] real businesses and build real solutions
[05:44] for businesses, that’s going to be also
[05:45] inside the community. But, anyways, so
[05:47] I’m inside my Visual Studio Code inside
[05:50] Cloud Code. What I’m going to do is come
[05:51] in the bottom here and I’m going to
[05:53] press {slash} g e o. And as you can see,
[05:56] all of these different {slash} commands,
[05:58] which are these skills, are available
[06:00] for me. I’m going to click on g e o
[06:02] audit. By the way, all of these details,
[06:05] this g e o audit and everything is
[06:06] inside this GitHub repo that I created.
[06:08] Again, this is completely free for you
[06:10] guys, available. Link is going to be in
[06:11] the description again. So, basically,
[06:13] this g e o tool consists of 12 different
[06:16] skills. As you can see, these are the
[06:18] different skills
[06:19] on the folders. Each folder has a
[06:21] skills.md file. So, this g e o audit, as
[06:24] I can click on skill.md, you’ll see
[06:26] exactly that all of these details are
[06:28] here and I’m going to talk about that a
[06:30] little bit. So, what I’m going to do is
[06:31] go ahead and run g e o audit on a
[06:33] popular website called calendly.com.
[06:36] Again, I’m sure a lot of people are
[06:37] familiar with that. So, I’m just going
[06:38] to press enter. What this is going to do
[06:40] is now Cloud Code is going to look into
[06:43] the g e o skills file inside this
[06:46] particular folder, this skills.md, this
[06:49] g e o audit. That’s because that’s
[06:50] exactly what we ran. And these are the
[06:52] These are instructions that I was
[06:54] talking about at the beginning of the
[06:55] video, where now we’re giving Cloud Code
[06:58] a specific instruction and making it
[07:00] become an expert in a specific area. In
[07:03] this case, it’s basically g e o audit.
[07:05] Again, g e o stands for generative
[07:08] engine optimization. It’s basically
[07:10] optimizing businesses for search in
[07:13] ChatGPT, Gemini, and Cloud software,
[07:15] right? So, what this is going to do is
[07:17] it’s going to run all of these analyses,
[07:19] as you can see. So, it’s going through
[07:20] and running these these multiple steps.
[07:22] And all of this is again coming from the
[07:25] instructions that’s inside this g e o
[07:28] audit skills.md file. So, the first
[07:30] phase is going to be discovery. It’s
[07:32] going to discover the homepage of the
[07:34] calendly.com. And then afterwards, it’s
[07:37] going to run this phase two, which is
[07:39] parallel sub-agent delegation. So, this
[07:42] skills has instructed Claude code to run
[07:45] five sub-agents simultaneously. Each
[07:48] agent is responsible for a different
[07:51] task, and that’s what the beauty of
[07:52] skills are. You can you can have a
[07:55] simple instruction, or you can tell
[07:57] Claude to run multiple different agents
[08:00] doing different things, right? So,
[08:01] imagine if you’re building a house. I
[08:03] gave this example before. If you’re
[08:05] building a house and you have a main
[08:07] contractor, then that contractor knows
[08:10] make sure that the house is built in an
[08:12] efficient way. They hire five different
[08:14] sub-contractors, right? Each
[08:15] sub-contractor will be in charge of
[08:17] doing something in particular. An
[08:19] electrician, HVAC contractor, one’s
[08:21] going to be building the foundation. So,
[08:24] that’s the best way to basically build
[08:26] that house in the most efficient way.
[08:27] That’s exactly what’s happening here.
[08:29] Behind the scene, Claude code is running
[08:31] all of these different sub-agents, and
[08:34] they’re delegating these different
[08:35] tasks. As you can see right here,
[08:37] everything is going through. I’m just
[08:38] going to fast-forward this so you see
[08:39] the result here. And as you can see, the
[08:41] result is running this GEO report. So,
[08:43] if I open this now, this is the output
[08:46] of this particular tool. And as you can
[08:48] see, it ran this GEO analysis report
[08:51] with the executive summary, score
[08:53] breakdown, right? Uh for different
[08:55] platforms like ChatGPT, Perplexity,
[08:57] Gemini, Bing, all of that stuff. By the
[08:59] way, I did a step-by-step on this. If
[09:01] you want to If you’re interested in
[09:02] checking it out, I’m going to put the
[09:03] link in the description of that video.
[09:05] But essentially, this runs this
[09:06] incredible service and runs these key
[09:09] findings for this particular company. It
[09:11] says, “Hey, Callanly has no Wikipedia
[09:12] article. This is a critical finding.”
[09:14] And at the bottom, it gives it also a
[09:16] prioritized action plan what they can do
[09:19] on a weekly basis to basically improve
[09:22] their AI sellability from ChatGPTs and
[09:25] Geminis of the world. Again, I walked
[09:27] through this step-by-step. If you’re
[09:28] interested in finding uh more details
[09:30] about this. But essentially, this is
[09:32] what this skills these skills files
[09:34] files are, right? This particular one,
[09:36] GEO audit, this is kind of like an
[09:38] entire comprehensive GEO Audit, but
[09:41] there’s multiple other skills inside it.
[09:43] So, for instance, if you just want to
[09:45] run a GEO site ability analysis for that
[09:47] particular website, you click on this or
[09:49] you run the GEO {slash} {dash} site
[09:52] ability instead of the GEO {dash} audit.
[09:55] And Claude Code is going to look at the
[09:56] skills.md file inside GEO {dash} site
[09:59] ability, and this is going to have a
[10:01] separate set of instructions. As you can
[10:02] see, this one doesn’t have any sub
[10:04] agents. So, if you run this, this has a
[10:07] particular set of instruction that tells
[10:09] Claude how to operate for that task that
[10:11] the user is asking for. So, that’s
[10:13] basically how you can run multiple
[10:15] different skills or or or give Claude
[10:18] Code uh multiple skills, and based on
[10:21] what you’re asking, it knows exactly
[10:23] what to run, when to run it. And again,
[10:25] Claude Code is smart enough to know that
[10:28] even if you give it plain English, just
[10:30] tell them, “Hey, I want to have a report
[10:33] generated for a GEO audit, right?” You
[10:35] don’t have to specifically run the slash
[10:37] commands. It will know exactly what
[10:39] you’re asking for. It will look into the
[10:42] skills files and see, “Okay, which one
[10:44] is going to be relevant for this
[10:46] particular ask the user is asking me
[10:48] for.” It’s going to pull that off the
[10:50] shelf, those particular instruction, and
[10:52] it’s going to follow through. And that’s
[10:53] what makes this such a powerful tool,
[10:55] and that’s why Anthropic is saying that
[10:58] the future of AI agents is actually
[11:00] building separate skills and giving
[11:03] Claude Code access to all the skills.
[11:05] And based on what the user is asking,
[11:07] then Claude Code is capable of pulling
[11:09] the exact skill that it needs in order
[11:11] to run the analysis or in order to
[11:14] satisfy the need of what the user is
[11:16] asking. Again, this was just a quick uh
[11:18] walk-through. I have done a step-by-step
[11:20] on this. You can check it out if you if
[11:22] you’re interested, but this hopefully
[11:23] give you an idea of what skills are, why
[11:26] this is the future. I’m going to be
[11:28] doing a lot more videos on this, so make
[11:29] sure you like and subscribe to the
[11:31] channel because I’ve got a lot more
[11:32] content coming up. I recently also did a
[11:34] marketing one. Uh so check that out if
[11:36] you’re interested as well. And again, if
[11:38] you’re interested in launching your AI
[11:39] agency and learning how to monetize
[11:41] this, check out the community. Link is
[11:43] going to be in the description. Thanks
[11:44] for watching and I’ll see you in the
[11:45] next one.

關係圖譜

反向連結