0:00 Today I want to show you a workflow that I've been using more and more recently and honestly it's become one of my favorite ways to build out AI skills. It combines Notebook LM's deep research feature with Cloud's skill system and that result is basically a way to turn raw research on any topic into a working AI agent that actually knows what it's doing. Here's a problem I keep running into. Whenever I wanted to build a cloud skill for something specific, whether that was SEO, copywriting, a specific industry, whatever, I either have to write the whole knowledge base myself, which takes forever, or I just prompt Claude with vague instructions and hope for the best. Neither of those is great. What I really needed was a way to get deep, structured knowledge on a topic fast and then feed that directly into a skill that Claude
0:46 can actually use. And that's exactly what this workflow does. Notebook LM does the research, Claude builds the skill, and then you got an agent that can be dropped into any chat or used inside Claude code that actually knows the domain. Let me walk you through this whole thing. Before we get into Notebook LM, let me quickly explain what a Claude scale actually is. Because if you haven't used them before, this is the piece that makes everything click. A scale in Claude is basically a structured knowledge file, usually a markdown document, that tells Claude how to behave in a specific context. It's not just a system prompt, it's more like a domain brief. You define what the task is, what good output looks like, what patterns to follow, what to avoid, and you can include examples, frameworks, step-by-step instructions, all of it. When you load a
1:33 skill into a project or point Claude code at a skill folder, Claude reads it and adapts his behavior accordingly. So if you have a skill for say, writing cold outreach emails, Claude isn't just guessing what that means, it's working from a defined playbook. That's the whole point. You're basically training Claude to be really good at one specific thing without having to explain yourself from scratch every single session. Now the question is, how do you build a good skill? That's where Notebook LM comes in. Notebook LM added Deep Research back in November. And if you haven't tried it yet, it's genuinely one of the more useful AI research tools available right now. And it's also free. Here's how it works. You go into Notebook LM, you go to the Sources panel, you select Web as your source type, and then you get two options. Fast Research,
2:19 which is a quick scan, and Deep Research, which is the one we want. You type in a question or topic and notebook LM literally goes and browses hundreds of websites on your behalf. It creates a research plan, refines a search as it goes, and then spits out a structured source grounded report in a few minutes. What makes it different from just Googling or using another AI search tool is that the output doesn't disappear. The report and all the sources it pulled get added directly into your notebook LM. So now you got this rich knowledge base you can keep curing, adding to, and cross-referencing. You can run multiple deep research queries on the same topic, layer in your own documents, and build up a really comprehensive picture of whatever niche or subject you're working on. For our purposes, building skills, this is incredibly useful because a
3:06 good skill needs real domain knowledge. It needs to understand how professionals in that space actually think and work, and deep research can pull that together for you in minutes instead of hours. Okay, so let me show you exactly how this works end to end. Step one. Pick your niche or topic. For this demo, I'm going to use something concrete. Let's say you want to build a cloud skill for B2B sales prospecting. You want a skill that can help cloud write targeted outreach, research prospects, identify pain points, that kind of thing. Step two, run deep research in Notebook LM. I'm going to go into Notebook LM, open up a new notebook, go to the sources panel, select web, and hit deep research. My question is something like, what are the most effective B2B sales prospecting strategies in 2026, including frameworks, outreach
3:55 patterns, and show how top sales team identify and qualify leads? You hit go, it runs in the background, and in a few minutes, you got a structured report with sources. Now, I'll usually run two or three research queries on the same topic to make sure I'm getting good coverage. Maybe one focused on frameworks and strategy and one on tools and workflows. one on common mistakes maybe, but each one pulls different sources and all of it gets layered into the same notebook. Before we continue, we just launched the Universe of AI newsletter. If you want to stay on top of AI news without having to hunt for it, link is in the description. Don't miss out. Step 3 is to synthesize a notebook LM. Now that I got all this research loaded in, if we take a look at this, we can see that this article over here at the top is a summary of all of the cited sources And
4:41 it kind of shows us what B2B sales looks like, especially in 2026 and how it is transitioning to AI intelligence and signal driven frameworks. So we have frameworks. We also have where it's heading and things like that. I can now start asking Notebook LM to help me organize it. I can ask it things like what are the core principles that come up across all these sources? What frameworks are mentioned most? What does good prospecting actually look like step by step? And Notebook LM pulls from everything in the notebook and gives me grounded, cited answers. This is the synthesis layer. I'm not actually writing the skill from scratch. I'm extracting structured insight from real research backed in by notebook.lm. So we can see our notebook.lm has imported all the sources. We have the summary document as well. We can see already that we also have top
5:27 sales methodologies over here. If I wanted to just learn about them, I can learn about them over here. But for our purposes, we want to take all this information and feed it to Claude, especially for the skill file. So I'm going to show you guys exactly how to do that. You have a couple of options. You can take this specific summary document, which summarizes all of your research from all the sources. But one thing you'll notice is that it's pretty vague. right? We have benchmarks, we have AI automation and things like that, not specific frameworks and everything. So you could do something simple like this. You could ask a specific questions like what are the specific frameworks or strategies and take that information for our purposes. Let's say I want to learn about how to write targeted outreach, research prospects, and identify pain points and build the skill around that. I'll ask that question and send it to notebooklm. then
6:15 it's going to go through all that research and give me a summary of how to do that specific skill so it's going to do that and it's going to send us an answer so you can see that we have an answer over here basically telling us that how to research prospects the five minute habit identifying pain points what we need to do and how to write targeted outreach so once you're in claude all you got to do is click on this sidebar click on that and then click on customize and then go to your skills and then click on creating one over here you can do create with claude or you can write skill instructions For first timers, try to do create with Cloud because it makes the whole process much easier. So we'll do that. You can see it gives you a prompt. Let's create a skill together using your skill creator skill. First, ask me what this skill should do. So instead of us starting from scratch here, where most people would when they're actually doing the skills, you now have research and methodologies that you can take
7:04 from Notebook LM and paste into Cloud instead of you typing from scratch. So we're going to take this and we're going to give it a name. Then I'm going to create a section called process of knowledge. And what I'm going to do over here is paste in all the information that we got from over there. Then you might also want to specify how you want the output to be for this purposes. Obviously I'm going to use this in multiple formats, but I can just tell it that in the output, use the knowledge base that we have given it over here to help generate work that aligns with the research. We want to do that. Let's say if you were kind of creating a skill for a specific writing style, you could do some deep research on their writing style, bring that in and tell it the same thing. And it will align your output to that. Or if you're trying to create an AI consultant, You can study all these frameworks and you can tell it to give you an output, specify the output like implementation plans, strategic takeaways, things
7:52 like that. This is very customizable. So you can play around with this and you can duplicate your skills and change your output format, or you don't have to specify how you want the output and can let Claude decide for yourself. But as I said, it's going to be dependent on the skill you're trying to create and the specific domain knowledge you're trying to understand as well. Right? So we'll do that. So I'm going to press enter. And what Claude is going to do is you're going to use that skill creator. you can see over here that skill that it has so right now we are in a skill section basically is using a skill file to create another skill file and then it's going to create that new skill which is going to be called outreach skill you can see they read that skill file but it's also asking some clarification question which is why sometimes building skills with claude is much easier because you know you might have missed something in your head that claude kind of flags for example it's asking three questions which output types called email
8:40 linkedin phone scripts or full multi-channel What inputs cloud should expect? How opinionated on structure, very rigid, flexible, or loose. For our purposes, I'm gonna select the full multi-channel sequence, which makes sense. Start off with email, then go to LinkedIn and then phone. So let's do that. And then the second question is, what should Claude need as input to use his skill? Prospect name from company, signals, triggers, your product service context, a target person or ICP. Let's do your product or service context and prospect name plus company. So I got to specify what I'm trying to sell and then the company and then Claude will do the rest. How opinionated should the skill be about structure? Very opinionated, always follow the three-part structure, context, value, ask. flexible, suggest the framework, but let context guide it. Just apply the knowledge. Don't enforce a rigid template. So this
9:26 is really good because Claude is saying, okay, you provided me with this framework. Do you want me to stick to it all the time? Do you want me to be flexible or just apply the knowledge and don't enforce a rigid template? I'm going to do the middle ground, which is kind of flexible. Suggest the framework, but let me decide what I want. So now it's going to build that skill based off those clarification questions. So now what you're seeing is our whole skill package over here. We can see apply this skill whenever the user wants to write cold outreach sales sequences, prospecting messages. You are an expert B2B sales strategist and copywriter. Your job is to help the user write outreach that is hyper relevant, signal driven and structured for response, not volume. We already have this AI agent or AI skill. That's going to help us a lot, save us a lot of time. And it says step one, assess what you have input available, prospect name, everything like that, action, skip,
10:16 As so, depending on what I give it, it also, it figures out what it needs to do. Step two research mode when needed leadership changes, funding rounds, strategic language, job posting, competitive moves, depending on the situation, signal density rule, pain point deduction to determine what the pain point is. I'll put right. Then this part is probably more relevant to us, which is outreach writing core principles under 125 words signal. First opening outcome over features one ass only no buzzword and three parts structure context value as so it has all that framework that we apply from notebook LM and then you can build that sequence. We can see day one channel. day one linkedin day two email day four phone call day seven linkedin so we have that you know length and everything like that so it has this sequence already built in for us so
11:02 we can say a bunch of time and we have subject lines and everything like that but you can see we now have access to this skill and we can just copy this to our skills and it's going to be available for us to deploy whenever we want so now what i'm going to do is actually put my skill to test so i've taken this snack company a snack company that i love this is like their product they sell Lotus seeds, so they have all these flavors. And what I've done is I've created a simple hypothetical situation, which is to build a cold email sequence for Airbloom, a snack company struggling with order delivery and logistics. Then I'm also gonna ask it to try to determine other pain points that they might have. So this is all hypothetical and I've given it a link. And all I have to do is now, if I just press this, I have access to all my skills and then I can go over here and then you have skill,
11:48 but we also have the outreach one. So we'll click on the one that we just created. we'll do that and then i'll just press enter so right now what it's doing is actually reading the outreach skill so now you don't have to repeat yourself multiple times or paste in frameworks all over again you have access to this wherever so now it's doing that And just based off of that, it read the skill and has given me everything I need. For example, we have these pain hypotheses, which are like order fulfillment gaps, maybe fragile product in transit, cold change or seasonal demand spikes, returns and refund ops, everything like that. And then what I have over here is a couple of emails that I can send to them. For example, a 14 day sequence. So 14 days worth of work is done now because of the skill I've built. Number one, day one LinkedIn, what to say, what to do. then an email chain. Then if I want to cold call them, what I can do
12:35 is call them over here, everything like that. Then I have another LinkedIn touch, another email point touch up with them and everything like that. So I have 14 days worth of work just laid out for me within a couple of seconds. And I also did a couple of personalization notes, like who to target. So a founder CEO is saying that it's likely a small team of one to three people. And right now it says that this person, Dave White appeared in the searches so I can use his contact to reach out to him, whatever. And then everything like that and then also I asked it to do a full research on it so this is the summary of what the product is and on the bonus pain point which I asked it to do beyond the stated delivery and logistical issues there are five other areas I found as well so all of this work was done for me within a couple of seconds and I have not only one day's worth of work done but 14 days worth of cold emails or messages set up and this was just
13:24 in the chat feature if you were to use the co-work feature in cloud as well you could do that you could also use it in the code feature because you have asked access to this skill wherever. In Coldwork, hypothetically, for example, you can access your skills just like you would access everything in your normal chat feature. But now you also have access to your calendar in Coldwork. You have access to your actual email for it to send drafts and everything like that. You also have access to automations and scheduled tasks and everything like that. So this skill, once you have built it from Notebook LM, you have access to wherever you go in Cloud. Even you can transfer the skill into other AI models like ChatGPT, or even a gemini as well but i personally like claude because of co-work and everything like that so i've chosen claude today as my demo so that's pretty much the full workflow you use notebook lm's deep research to pull structured expert knowledge on any domain you want
14:13 then you take that research into claude and build a scale file that turns that knowledge into repeatable agent behavior and you end up with claude that doesn't just kind of know your domain it operates from a real knowledge base and gets better every time you update it. The library of skills you build up over time becomes genuinely valuable and it's the difference between using Claude as a generic assistant and using it as a specialized agent that knows your worth. If you want to see me apply this to a specific domain, drop it in the comments, I'll pick a few and maybe I'll do a dedicated walkthrough for that as well, but if you found this helpful, make sure you guys like the video and subscribe to the channel. But that's it for today's video, make sure you guys are subscribed to the channel, follow our new newsletter as well at universeofai.beehive.com, as well subscribe to the main channel, World of
14:58 AI, and support us on X by following the Universe of AIZ as well. Until then, I'll see you guys in the next video.
