Transcription
So, you might have set up your OpenClaw, your Hermes agent, or maybe even some derivative of my Claw Claw system, and it's running, but you've yet to derive any form of actual business value from it. Because the truth is, a lot of creators, including myself, show you the final finished product of what the ideal agentic OS system looks like. But in reality, the magic is in all the grunt work that happens behind the curtain. And this is the grunt work that isn't very sexy to show on YouTube, which is why very few actually talk about it.
So, this video is meant to help you get your house in order so you can take advantage of the AI opportunity and properly use these kinds of systems and finally gain leverage with them. So, if you have an idle or underutilized personal assistant or agentic OS system, then you're going to want to watch this till the very end because I'm first going to walk you through the mental model and mindset shift that you need to focus on the data before you worry about any form of dashboards or pretty features. And on top of that, I'm going to give you the exact skill we give all of our clients that we consult for to help them and to help you self-serve to make sense of all the files, integrations, APIs, and CLIs that you might need to fully get the leverage that you want. And I'll finally walk you through some tactical tips that no one else is focusing on, probably because they've only set it up for themselves, not other businesses.
All right, so big picture, this is the full cross-section of an Agentic operating system. And typically you are staring at the very top on YouTube. And the reason being is typically when we talk about data engineering, data principles, data readiness, this is where people fall off and their eyes glaze over. But to make it work, you have to make your data work as well. Now beneath this beautiful layer, we have skills where you have project level skills and global skills. Project are specific to a folder. Global apply to every single project irrespectively. And then we have one of the most underutilized and misunderstood features in something like cloud code which are hooks. And irrespective of who I speak to, whether it's a business owner or someone who works in a business, hooks are this weird innocuous thing that no one touches, even though they're actually very straightforward. They're very reliable and they can come in very handy. And last but not least, we have the Claude MD, which to many people is interpreted as a diary, but it's really an air traffic control document that you can create a whole empire of Claude MDs to give the right information to Claude Code at the right time. And on the right hand side, we have all of your integrations. And if you use something like Claude Co-work, then this is pretty simple to set up. It's a series of clicks. But if you're using cloud code, you have to be more intentional and think, do I need an API? Do I need a command line interface? Do I just need a skill? Do I need all of them? The answer might really depend on your specific circumstance. But the real point I want to drive home is that there are millions of people who have set up their open clauses and only a handful of them have derived real economic ROI from doing so because you're building a beautiful home with fantastic agents on top of burning rubble. And that rubble comes down to messy files, gigabytes of completely dated or useless information, no form of a business in a box operating system before this AI stuff even existed. So if you layer on sparkly slop on top of a foundation of slop, all you'll have is compounded slop.
Now, let me give you a tangible example so we can put some color to the problem. Let's say you are a business owner, a solopreneur, someone who works in a business and does reporting. And all you want to do is ask a very basic question. Let's say you're in e-commerce and you want to track the last six months of sales and you want Claude Code and your team of agents to derive all the patterns to lead you to success and tell you when the next event should be, what the discount should be, what your perfect avatar should be based on these six months of information. On the face of it, these models are advertised as being the silver bullet, meaning you can feed everything to this oracle and it will figure it out and synthesize it. But if you send off your agents to go pull six months of data, what will happen is the majority of your context window in that conversation will be bloated with a series of JSON and metadata that's invisible to you but necessary to pull if you're pulling this information. So when it comes to the actual valuable part, which is the synthesis analysis, and the real brain power that you're seeking, it falls flat. And the reason why is you're bloating it with things that don't matter. So one of the many ways that you can solve this is to put a lot of the core data, the KPIs that need the analysis part on a silver platter. So one of the many ways you could solve this is through summary files or summary tables where you distill the core information, the core KPIs from each one of those months. So if you ask your agent to go do a deep level analysis, it's actually spending time analyzing, not retrieving information and then analyzing. Because you can imagine if an agent is spending 80% of its session pulling the information just to be able to start analyzing it, you'll only get that last 20% of the session, which is usually the part where hallucinations happen, slowness happen, and weird behavior happens. And now you're dedicating it to do the very task you wanted to.
Trying to execute this more tactically is the very reason why I created my skill/splatter. And the way it works is this. Whether you've never started on your AI journey or you've already drafted your operating system and you want to optimize it, it will take a look at all the data you have on your computer. So if you have some form of integrations or you need it to talk to the cloud, you would need to make that as a prerequisite for this to do the analysis it needs to. Once it has an understanding and a snapshot of your data, your dotcloud rules, your skills, your claude MDs, then it will try to help you distill what could you summarize. What could you create summary tables or summary information to put the prerequisite information on a silver platter so that the agents can focus on the synthesis, not the pulling. And the end goal is eventually you have enough agents that are very focused and you give them all what's called a critical path. Meaning they don't have to go and figure out every single time where to look for the data and how to pull the data and how to analyze the data and which anomalies to regard and disregard. They have a good step-by-step process, an SOP of exactly what the expected behavior is. And on top of just having clarity, the skill will also help output the following HTML page. This will walk through your specific systems and files and segment it into three sections. A pantry, which is where all of your core services and databases live. Your prep table, what you can actually do with this information, and your plate, what you can do to actually tactically put this data to work. And if you go through the other tabs here, we have a data flow tab where out of the box, the skill will break down all the relationships between your data. So you can see how you can interact with it, where you can interact with it, and how you can have some cohesion between all your systems so they can actually speak to each other. And if you need suggestions on what you can build with the current data that you have, you can have that tab autogenerated for you. And the final thing is a 30-day plan. So this all comes out of the box from the slash command that I'm about to show you in a few minutes.
By the way, if you're finding this content helpful and you want to take your Agentic OS system and your Claude Code skills as well as your AI skills in general to the next level, then you're going to want to check out the first link in description below. I keep adding practical and applicable lessons to our living Claude Code course, and we keep improving day by day our existing Claude Claw operating system. So, if you want the best of both worlds in one place, then you'll want to check it out. All right, back to the video.
And like I said before, I've spoken about hooks on this channel multiple times. And typically those videos do the worst because hooks still break people's brains even though fundamentally a hook is an event that will always fire at one of 18 plus events that happen in cloud code. So if you start a brand new session, you can make something happen always in the beginning of that session. Some injection of memory. If you have Obsidian, you can make sure that if you're in a specific folder and you have Obsidian files related to litigations, court cases, anything around revenue in your business, you could always throw that in along with the context from your CloudMD. Similarly, if you have longer conversations with your agents and you find that when the conversation compacts, you lose a lot of core information and you have to repeat yourself, you could use what's called a post-compaction hook to inject that core information about you, your business, what you do, who you are to make sure they have a much richer conversation. And the best part is you can literally ask Claude code through the Claude Code Guide sub-agent to create these hooks for you. You just have to ask and you just have to know that you can ask. And when it comes to things like organizing your agents, there's a reason why you typically see the configuration here on the right hand side where you have an orchestrator or a chief of staff managing the other agents. The reason being is you can load one agent fully with the roles, responsibilities, and the scopes of the sub-agents. This will increase the likelihood that you don't end up in this scenario where you are cold starting conversations with your agent teams having a mismatch or overlap of which agent should fire at the right time.
Now, instead of just describing it, let's put this slash command to work. So if you go into a terminal and you use my skills silver platter, what it's going to do is if you have infrastructure that exists, meaning you have a cloudmd file, you have skills, you have rules, it will do a quick audit lasting from 10 to 50 seconds and give you a quick TLDDR of what the current state of your system looks like. And once it does the quick audit, it will say, "Hey, welcome." Walk you through the mental model of what we're doing here. And then it will say, "You've already started building." In my case, I see a claude MD, 38 skills, 16 rules, and a well-structured output pipeline. And then you can pick one of two paths. Obviously, you can take the skill and edit it to make a medium path if you need to. One is very thorough around 30 minutes of interview questions back and forth. This is really designed for someone who's starting from scratch or wants to restart depending on where you are in that process. In this case, I will pick the fast-track option just for the sake of the video. And then I will say, let's pick the fast-track option. And please make sure you don't use any real client names or people names or deliverable names. Keep it all anonymized through this process. So we'll send this up. And I just said that to make sure that you can't see real client names and emails. And this will come back with a full breakdown of questions for you. And then it's designed to pepper you with questions. So in this case, it starts off confirming what I do for a living and how I make money. And in this case, I just walk through that we are a consultancy. We have YouTube stuff. We have a community. We have high-ticket consulting. Then it asks me questions like where do I publish and what products do I have? It's trying to get a sense for what different APIs and CLIs could I use for my AIOS. Then asks me questions about my audience size, where I have video transcripts for meetings, workshop materials, if I use something like Obsidian, and then it asks me questions about my paid offers, where I post or host rather my email list, my audience research, how I pull my YouTube transcripts, a few questions about the day-to-day, about the business. Then eventually, it goes to start creating the data map as you can see here. So if we take a look at the output, it recommends a series of different things that we can look to automate. So it's very tactical. So I can say, let's do all in order. And from that vague prompt, it will hold your hand to execute each one of these steps. So this first one allows you to drag and drop your raw PDFs, spreadsheets, email exports into this data folder. Again, you can go back and forth and customize this as you wish. Then it will help you build your morning brief. Then it will tell you exactly if you were to put an or AI orchestrator for other agents what that would look like. And then as you scroll to the very bottom, it comes up with a series of ideas for slash commands. So in this case, maybe weekly triage for email, slash weekly content for my YouTube funnel, etc. And the last question it asks you on the fast track is where do you actually want to read these briefs? Do you read it in Gmail? Do you read it on your iPad through Slack, Telegram? And the reason it's asking you this is to help you decide if you set up your own agentic OS, do you want to interact with it from Telegram, from Slack, somewhere else? It's helping you optimize where your end usage is going to be. And once it has that, it starts putting together your full data map. It creates the JSON and associated HTML file. So this is all out of the box. And by the end of this, you will have something like this. So you have your full audit and then you can see all of my data sources. So YouTube Studio, you have School, you have Gumroad, you have Kit, all of these things including my Obsidian. You can click on them, better understand exactly the pros and cons if it has anything like a skill that you can install, and you can obviously ask Claude code or similar to go and search if you want. And then on top of that, you have the prep table on how we could add some automation to looking at things like monetization. And then additional things here like an inbox brief like we saw, a finance brief every Monday. And it walks you through what that would look like. And then like we saw before, we have the full data flow where I would end up reading the majority of my data. And then if we go to the very top, here are additional ideas specific for me. So, one example is Friday content brief with three angle candidates, each backed by data. And if you scroll down, it walks you through what the chain could look like. You have a cron job that spins up maybe every Friday at 5:00 p.m. and then you have a chief marketing officer bot that goes and clusters the signals, etc. So not only will it help you create the mental model of how to approach these problems but also it will give you the inspiration on what can be done that's very tactical.
And again, from a big picture perspective, we really have four layers. This number one layer is one of the most important where you really establish your identity, your cloud MD, any specific rules if you have GDPR or SOC 2 compliance stuff you have to abide with. And then the second layer is knowledge. So knowledge can live in the cloud, it can live in a drive, it can live in the form of skills that can access them or MCPs or both. And then you have the workers where you would set up your crystallized or in this case materialized meaning it comes out of the box. You actually have an agents MD file with the name of the agent that you can call on every time. And then this is where you establish what does it look like to have your own AI workforce. You don't want to have agents for the sake of having agents. You want to hire agents like someone would hire employees at a bootstrapped company. You don't hire five people all at once. You only hire an agent as you see that there's so much burden on one particular agent that it makes sense to naturally split it off into a separate subject matter expert. And once you've cleaned all the skeletons in your closet, you can ascend to layer four, where you can start layering on things like automations, like the injections of hooks to really enrich your conversations and keep things as deterministic as possible while using a very chaotic piece of software.
Now, to really drive the point home, let's take three hypothetical scenarios of three completely different avatars on how they would implement this exact same methodology. Let's say we have persona numero uno named Marco and he is a solo founder of Slab House. He basically does live streams for mystery boxes and he has multiple tiers. There's a very small tier, there's a whale tier, and one in between. And his stack is a combination of Shopify, we have Facebook ads, we have TikTok, we have Twitch for the live streams, we have Facebook Live, we have PDFs and a CSV empire which I know many clients have today. He struggles with things like P&L sheets, inventory, sales data, and he does this for 3 hours every single Monday. There must be a better way. Right? Right. We can get to a place where he can get a morning brief of his financials. As long as all of this data lives in a place and lives in a way that it's easy to be queried. So assuming that these are disparate CSVs that he's pulling directly from QuickBooks and manually analyzing, we could actually pull directly from the QuickBooks API, have Cloud Code create summary tables where depending on the metrics that matter, it will use Python deterministically so we're not risking an AI agent hallucinating. It will pre-aggregate things like revenue, cost of goods sold, all those fancy metrics and put it into very easy-to-read summary tables. Then when we have a cron job or we have something like an Open Claw, a Clawed Claw, a Hermes agent, whatever, it will be able to easily read those numbers and report on them with very minimal chance of hallucination because we did all the hard work for it. We did the 80 and we were now bringing the agent to actually look at the information and communicate trends, go through past versions of our conversations to look for patterns. And this is what his data map would look like where if you scroll down, these are the core inputs that we saw before. And then notice here there's a legend that says you have this either this is missing or it would be a quick win. So if you scroll to the bottom, one example of something that would be a quick win is a pre-stream prep brief where maybe there's a slash command. And here's exactly how Claude Code could set it up for him. And the result is that his simple questions would now become a chain of actions where he could ask a question like, "Is whale tier worth the cost of a guaranteed pull?" And then it would go to the orchestrator. It would go and talk to the CFO bot, which makes sense given we have a big emphasis on finances. And then it could talk to the chief marketing officer bot where it would look at the customer voice, analyze different trends, and see how it would be received by his audience.
Now, let's pivot to Sally, who's an associate at a small boutique law firm. And what she works on are small-cap merger and acquisition deals, four to five new matters a month. A matter is like a case in the law world. And then this is her stack. She has Outlook, she has Bill for Time, which is their billing system. She has unfortunately Microsoft Copilot, but for Claude Code, they use Amazon Bedrock because everything needs to be well-controlled and well-contained because it is very confidential client data. And just like other law firms, it's very PDF doc and email intensive. And let's say that the current state is that she spends 3 hours to go through the matters, scaffold all the ideas, review and write the briefs, etc. We could get to the point where we have a slash command called slashcase launch, where it would execute a series of processes that she could outline in an SOP that she does over and over again because the real leverage in an agentic system is that you create a skill and you keep iterating on that skill non-stop. So a skill is an infinite game. It's not a finite game. You don't finish making a skill. You start a skill and you keep improving it over time by going through a conversation, going through all the wrong turns and eventually getting to the right turn and then saying, "Based on all our conversation, I want you to map out the perfect critical path that would prevent us from going down the wrong avenues again." And your eyes might glaze over as I describe this, but I'm telling you, this is the difference between someone who works at 80% capacity versus 100% capacity. And then in terms of her report, if you click on what we'll build and we scroll down and we go to new matter scaffolding, it gives you the idea of creating either a skill, adding a hook as an auto-trigger, but basically creating this specific chain where she can drop a brand new PDF as a new client intake, maybe she runs a command that's called /new matter. And this is exactly how it would launch.
And last but not least, we pivot to Dr. Santa Anoir who's working in the dermatology domain. She does have to abide also by healthcare data privacy laws. And then she deals with things like Athena Health, which is a specific CRM that's very common in the healthcare domain. And then she has to deal with intake forms, EHRs, which is the basically CRM. And then she has to process all kinds of biopsies. In her world, creating the Agentic OS is even more complicated because she has to really bifurcate where all this data lives. We have data at the clinical level and then we have it at the billing level. Given the constraints of this domain, you have to be very mindful about isolating each and every data set and handling in different ways. So unlike a solopreneur, she can't just tell Claude Code, "Don't make mistakes." She might need to fully isolate each one of these data sets and create a unique set of hooks, Claude MDs for each and every domain. So her data flow might look very different from the average person where you go from this EHR system to all of these different prep tables that process the data, and you'd have to be very mindful about how the data is handled in transit. And obviously, you get the idea now. So if we go to what we could build and we scroll down here, you could see that we could create these skills like summarize biopsy PDF, ingest Athena encounters, care coordinator orchestrator, etc. And these are examples of different rules and hooks that could be helpful. One really important thing for her setup are hooks because she can't risk any form of data or names being sent to Anthropic servers depending on the situation. At the end of the day, context is king in this domain. So, the more context you can provide and the more cleanly you can provide it, the higher the likelihood that when you have an agentic system, if you want to take your agents on the go like I do while I'm traveling and have them actually drive real business value, all the hard work is in everything that happens behind the curtain.
So, hopefully this really demystifies and simplifies what you would have to do to take your Agentic OS setup to the next level. And like I said, I'm making the skill I showed you in the video available to you in the second link in the description below. And naturally, if you want to go much deeper on Claude Code and learn while you build and actually perfect your Agentic OS system, then you want to make sure you check out the first link in the description below as we always have up-and-coming courses in my living Claude Code course. And on top of that, you'd be able to take advantage of my Claude Claw business operating system, so you can do the dirty work behind the scenes and then layer on top a very powerful Agentic system. For the rest of you, if you found this helpful and illuminating, I would super appreciate a like and a comment on the video.