Loop Engineering, Claude Code /goal, and Agent Routines

So today I'm going to talk about loop engineering, which is the buzzword right now. And in this tutorial, we're going to build your first autonomous agent using the concepts of loop engineering and using specific tools such as Claude Kold. using cloud code, the goal command and routines. And this tutorial is really for AI builders who have heard a lot about autonomous agents, but you don't have hands on experience yet building your first one. So in the first part of this talk, I'm going to talk about what is loop engineering, like literally, let's just define it and understand the key concepts behind it. And then let's go ahead and build our first autonomous agents using these cloud code tools, such as slash goal and routines. And by the way, you have full permission to repurpose all of my education. So today's newsletter is literally thousands of dollars worth of training. I know this because I see a lot of AI content. Most of it's very shallow. Everything today is like from real world practical experience supporting thousands of customers with my products. And honestly, many of you will fall off on this training, okay? You're going to need to go to this tutorial and follow through after, but it's 100,000% worth it. But... Yeah. Oh, my husband just locked it. Wait, what? He just walked into my room and stood in the corner. Okay, so I'm just gonna I'm just saying that because I know this is gonna be a lot today So it's okay if like you fall off somewhere. That's why I made a companion newsletter and I will send this out immediately after and I just want to reiterate you have full permission to repurpose all of this education content, okay So yeah, so here's the agenda. So today we're going to talk about what loop engineering actually means, and then we're going to implement it for real using a clod code goal command. I'll teach you exactly how to write the perfect slash goal prompt and how to make sure it doesn't blow up everything on your computer. And then we're going to build our first autonomous agents by combining the concepts from loop engineering into a clod code routine. You do not need any technical background for this. You don't need to know, even though we're using clod code, you do not. need to know how to code. However, the only prerequisite here is setting up cloud code. So if you haven't done that, click on this link or go to my YouTube and search cloud code tutorial for beginners so that you can set it up. Okay. Okay. So first, let's just talk about what loop engineering means. Like what problem is it solving? So the typical way people do things is they prompt AI, it could be chat, it could be Claude, they type, they read the answer from AI, and they type again, they read the answer from AI, you type again. So in that case, you are actually part of the loop. The loop is you and AI. You are the one checking AI's work. you're deciding what the next step should be, and then you're telling AI. So you are the loop. Loop engineering, all it really means is taking you out of that loop and building a little system that does the prompting for you, meaning it checks the AI's work, and then it decides what the next step should be. And when I say it, I mean another AI agent. So loop engineering is really about building a system that prompts your AI on a schedule, and against the goal. Okay, you need to have a goal, otherwise the AI doesn't know if it should stop. But once you build a system like this, the beauty is you don't have to type every single prompt yourself. You don't have to review every single output yourself. So all we mean by a loop is like, do a step, check the result. Did we accomplish the goal? If not, here's what you should do next. If we did accomplish the goal, then we can stop. Okay, so there have been different ways to do this. Like, for example, I often will append this to the end of a complex prompt. Don't stop until dot dot dot. But now with Cloud Code goals and routines, there are more formalized ways that you can do this. Okay, so here, yeah, I'll talk about the loop has six parts. Okay, here, I have pretty visuals now. Okay. Okay, so just to put, and by the way, this was all AI generated visuals, so super cool. Here, I'll play it for people who didn't see it the first time. This is all AI edited, so this is AI generated as well. So again, like the way 99.99% of people who are using AI right now, you are in the loop. Okay, you and AI are talking back and forth. You say, fix this bug. ChatGPT says, okay. I fixed it. Then you go take a look. You review the work. And you're like, oh no, this still doesn't work. Here's why. Then chat's like, oh, okay, you're right. Let me take a look again. All right, like that is the loop. It's just you and the AI. are in the loop together. So loop engineering, we're trying to remove you from this loop and replace you with another AI agent. So now we're going to have two AI agents, one doing the work and the other one doing what you would have done, which is review the work and then decide what the next step should be, what the next prompt should be. Now, the concept of loop engineering that is being hyped right now, it's really the same thing people have been doing, or at least advanced people have. doing so when you type a prompt like don't stop until this condition is met that is loop engineering essentially now like the tooling has improved so that this is now much better okay so we're gonna go through each of these one by one so the very first i should back up there's six key concepts in what people are calling loop engineering today so the first one is automations let me just go to the right visual okay here So all we mean by automations is like there's some kind of schedule and then your AI agent runs and it completes certain tasks and it's running by itself. So instead of you sitting there every time and telling AI, here's what you should do, the loop is running on its own. It's running by itself on a schedule or you can trigger it with an external event. So that's automation. Okay. The second part of a loop is, let me go forward here. Da da da. Okay, here. The second part of the loop is called work trees. And this the concept here is when you have many different AI agents, and they're all working on the same project, they can easily conflict. So what you want to do is give each agent its own lane, you can think of it as like lanes on a highway, which AI made this visual I really liked like lanes on a highway. Okay, if you have like six agents all working on the same project, changing the same files, it's very likely to end up in a complete mess. Like this agent's going to change this thing, but this other agent was working on the same thing and now they're confused. This third agent was reading and analyzing the thing and then it changed because of some other agent editing it. What you want is different, what they call work trees for each agent. Now, just a caveat, you don't actually have to implement this yourself. It's now part of cloud code and tooling. Okay. But conceptually, this is a really important concept because like once you're orchestrating multiple agents you want each one to have its own lane its own work tree so it can safely make changes without disrupting the work of other agents simultaneously okay now concept number three behind loop engineering is skills many of you are probably familiar with skills okay but for those who aren't it's you can think of them as a playbook you write the way like let's say you reply to customers support tickets a certain way, you want a playbook that the AI agent can follow. So it can consistently reply to customer support tickets the exact same way that you would. So a skill MD file just formalizes whatever your playbook is. So for example, how you want to name things, the conventions you want the AI agent to follow, the mistakes that it should never make, any constraints or guardrails. You would include all of this in your skill MD file or your playbook. Okay, so you teach AI once and it will follow this playbook skill every single time. Okay, key concept number four in loop engineering. Let me sort here. Key concept number four is connectors. You can think of this as AI being able to use the tools that you use every single day, such as Gmail, Slack, Intercom, etc. So here is a good visual. Okay. So instead of AI just yapping back and forth, like here's the fix that you should implement, connectors allow your agents to go implement the fix for you. Like it'll fetch your code base, find the bug, try to fix it, add helpful comments, and check it back into your code base versus AI that's just yapping and telling you what to do. Okay, so when you connect your AI agent to connectors, it can basically use all of the tools that you use could be gmail could be your issue tracker uh jira github linear could be your communication systems like slack could be your crm like salesforce hubspot content creation tools like canva higgs field potato etc whatever tools you use now your ai agents can use them too this is already a really huge unlock so like personally i just try to get people to this step you know like if you can get to this step you're already ahead of 99% of people using AI. Like, I'm a very happy teacher. But what we're going to talk about today is like beyond this. I'm just kind of reiterating what are the key concepts in loop engineering. And by the way, for many of you who have been using AI for a long time, who have joined my live streams for a long time, you should already be familiar with most of these concepts. And that's kind of what I'm trying to emphasize. Like, loop engineering can... as a term is brand new, but as a concept, like we were all doing different pieces of it. And it's kind of blown up on Twitter because now we have a term that we can all rally around and say, oh yeah, this is what I've been doing. This is how I've been thinking about building my agents. Okay. So blah, blah, blah. Oh yeah, that was a good visual. Okay, so concept number five in loop engineering is sub-agents. This one is really important. So the easiest way to think about it is you have two agents, the one who's actually doing the work, the maker. and then a second agent, the checker. And the reason why you want to have two separate agents is because the first agent is just biased and will lie to you. You know, when you're like, fix this bug, and then your AI is like, yeah, it's done, everything works great, but it actually just deleted all of your tests. Yeah, you don't want the agent doing the work to be the one checking the work because it will often tell you that it's done even though it's not done. So the key concept here is we're just separating the tasks. One agent will do all of the work and the second agent will independently review it. If the checker says, hey, this isn't good enough or this isn't ready yet, it's going to tell the maker, here's why and here's what I think you should do next. Right, this is very intuitive if you just think about it as like two people interacting. One person making stuff, the other person reviewing the work. Like a writer and an editor. I think I have a visual for that actually. Oh yeah, nice. So yeah, like, so a writer's writing and an editor is like a second set of eyes and they are paid to be brutally honest. Okay. And it's this like interaction back and forth that leads to amazing writing. It's that collaboration between the creative person who's doing all of this writing, but also the editor who's taking a fresh look at it and giving really critical feedback. Like, is it done? Is it ready? Is it high quality? So a lot of people don't do this. They will use the same. AI to review the work done and that's why you get a lot of issues like it said it was done but it really wasn't it was super broken but why did it lie to me because it's reviewing its own work okay so that's concept number okay we're blasting through these okay six So the final concept in loop engineering is memory. And you can think of this as like a shared notebook. Okay, so that your AI agents on each run has context on what's been done, what else needs to be done, what should I do next? Now, practically, when you hook up Cloud Code to your GitHub repo or your GitHub project, that serves as a sort of memory. Because like within that project, for example, in my social media posting, project my agent tracks like all of the posts it's made like it has a little log file that tracks that so practically here you don't have to do anything if you've already hooked up Claude to your GitHub repo but the reason this is important is without some kind of shared notebook across all of your agents every run would start from zero. Like it really wouldn't know what the previous agent has already accomplished, what blockers they're running into, and what it should do next. So memory is an important part of loop engineering, but practically you can achieve this just by connecting your Cloud Code project to a GitHub repo so that your agents can write to that project, their status updates, what they've done, what else they're stuck on, and what they should do next. Okay. Oh, yeah, it's like a whiteboard. Okay. Oh, this is a good one. Okay, cool. And by the way, for those who joined late, what I'm showing on the screen is a YouTube video I'm going to publish later. And it was fully edited by Claude and Hyperframes, which is a new free open source repo. And it's pretty cool. Like, it's very cute and everything. All I did, I just filmed myself talking, and then it edited it and added all of these visuals and interactive graphics. They're not moving right now because I just, I don't. I don't want too much motion so I can explain stuff, but all of these graphics are animated and stuff. Here I can show this. Okay. And then, oh, that's the whiteboard. It even cut out my face here. Oh yeah, here. I was just showing the animations. Okay, cool. Okay, so yeah, let me just summarize. So in summary, loop engineering is all about removing you, the human, from the loop so that an AI agent has a goal, it knows what to do, It tries to accomplish the goal. It reflects on its own work or a separate agent, a checker reviews its work. And then it decides, like, if the goal has not been met, here's what I should do next. That is what Boris, the founder of Claude Code, means when he says he doesn't prompt Claude anymore. His whole job is designing these loops. Because instead of like, I ask Claude, do this. I review the work. and then I tell it what to do next. Instead of you being the one doing that, you can have AI taking your role doing that, reviewing the work, and then prompting what should the next step be. So here's the whole loop that we're going to build, by the way. uh they basically have some kind of trigger like a schedule when your agent runs it'll have a shared notebook or memory which is typically your github project if you have multiple agents working simultaneously each one will have its own work tree its own lane so it doesn't collide with all the other agents. You'll have one agent that is predominantly doing the work, the maker. Another agent, usually faster and smaller, such as a cheaper model like Haiku, checking and reviewing the work independently. If the goal has not been met, the checker says, hey, go back, do this instead. Like, it'll give feedback to the maker in terms of what steps should be taken next. That's what... people mean when they say let AI prompt itself. The checker is actually providing feedback and in context for what the next prompt should be for the maker agent who's responsible for doing the work. And then when it's when the work is done, you can open a PR, for example, you can email it to yourself, you can send a Slack message to yourself exam. So okay, cool. So I think that's it that I want to talk about here. Yeah. Next turn. Yeah, okay, cool. Okay, do we have any questions on that? That was a cool video. Yeah. Okay, if you guys have any questions on that, let me know. Just drop it in the chats. Are you actually using loops like this for Blotato features? Yes. So that's what my new AI agent is. So I dropped a new AI agent in Blotato that... uses literally exactly this loop. And it does not use any bloated agent frameworks or anything like that. It's really just built from scratch without any agent frameworks. Can I use codecs to loop Cloud Code? You can, but I'll show you today. You can also do it entirely within Cloud Code. Could we do... A quick summary. Okay. Oh, here. I put the summary here. I'm going to send out this thing immediately after. Okay, so here were the six parts we talked about. Number one is, oops, sorry. Number one is automation. So some kind of trigger or timer that starts the loop on its own so that you don't have to be the one typing to start the work. Okay, we're going to cover that. in the last section, cloud code routines. Second concept number two is separating the work areas. So two AI agents do not collide with each other. Practically speaking, you normally don't have to set this up manually, and most people won't even have this need, to be honest with you. Like, you'll just have one AI agent at a time working on your project. It's a lot to manage multiple work streams, so I would consider that very advanced. But concept number three, skills. So having your loop. access repeatable playbooks for the different tasks you should do. Many of you are familiar probably with skills and connectors already. So the concept here is just that your loop has access to skills as well as connectors. Connectors allow your agent to use the tools that you use. So Gmail, Slack, HubSpot, whatever you use, you can plug in. And sub-agents, the concept here is one agent to do the work and a different agent to independently check the work so that your main agent is not grading its own homework. And then the last concept is memory. So that could be a notes file that lives outside of the chat. It's oftentimes a GitHub project. with text-based markdown files within the project where your agents can log information, etc. A lot of people now like to do a lot of markdown in Obsidian and use that as their second brain memory. So that'll just stop there and summarize instead of rambling. So, okay, so those were the six parts of loop engineering. And if most of those look familiar to you, that's good. They should. Right, loop engineering as a term is not a new concept in the sense that like people have been doing this, but it's blown up on Twitter recently because like someone really put a name around it and helped to kind of like organize a lot of the thinking that has been done around this. So if you're all, you're probably already doing some of these, for example, if you have a cloud cowork schedule that checks your Gmail every morning and surfaces. emails that are urgent to you and labels or archives the rest that is considered loop engineering it may not have all of the parts okay but conceptually you've gotten a lot of things there already okay so those were the key concepts yeah I will caveat this like it sounds simple but actually the hardest part is defining the checker like Let's take a social media post for example. How do you define what a good post is? Like it seems like your task is like write a viral LinkedIn post, okay? But like how does the checker define what a great viral post is? And so as we'll walk through the examples today, eventually you'll notice that actually defining what good and done looks like is the hardest part of creating the loop. Oftentimes people already know the task they want done, right? Like clean this up or write this thing or plan this out. But it's hard to define what good looks like, what done looks like. But that's okay. That means you're doing it right. So, okay. Okay, so now we're going to dive into the actual practical part of this tutorial. So make sure you have clod code open. Uh, let me actually open it now on the right side. Oops, that's okay. I'm just trying to get the font size right, like, so you guys can read the font. Can you guys still kind of see the font, the stuff on the right side? Hopefully. I know it's really tough on vertical streaming platforms. I am on horizontal streaming platforms as well, if you find it really hard to read. You can use cursor and other tools, but they won't have slash commands like slash goal, I think. Honestly, I haven't used cursor for real in like a year. And then I'm not sure about like routines either. Before Cloud Code routines, I just used GitHub Actions, right? So that's another option if you're GitHub based. Okay, cool. Can you guys see my screen? No, it's actually the opposite. The checker is the small, fast model. So the default checker in Cloud Code is Haiku. You can change it so it uses a different model, but typically you want your primary worker to be a more advanced model, and then your checker to be a smaller, faster model. Okay. Okay, cool. So yeah, let's get started. So the very first thing we're going to do is just think about... a task you repeat every week. So when do you start the task? Like literally fill this out. Starts when? Is it a certain time of day? Is it when an email arrives? And then define what done looks like. So for example, in the morning at 8am, I want to go through all my emails and label them accordingly or archive them and And let's say surface the three most urgent ones to my attention by DMing me on Slack, something like that. And then here's the check, how the AI proves it's done. That would mean like every email, it has been labeled or archived. Okay. So think about this for one task that you repeat every single week. And by the way, I recommend business owners in particular do this once a week. So this is an aside. This should not be in the final video. But I have a habit tracker app that I always use when I develop new habits. And one of them that I've used now for a while is always to automate, delegate, or subtract something I'm repeating every single week. So if there's something I'm doing every single week, I either want to automate it with the help of AI. delegate it to someone else or subtract it from my life. Just stop doing it altogether. If you do not have that practice in place on a weekly basis, you'll often end up like just drowned in lots of random tasks. So, but so anyway, that was an aside. But okay, so I'm just giving people time to think about this one sentence they repeat, or this one task they repeat every week. Okay, now what we're going to do is use the clod code slash goal command and make sure you have clod code set up in terminal. Okay, so I'm just scrolling down here to where the prompt will be. And similar to the concepts we talked about before, so you give Claude your goal, like what is the goal condition? It will keep working on its own until the goal is completed. So how does that work? So for those of you who are technical and curious how this actually works, basically slash goal is a wrapper around the stop post hook notification. So you know post hook notifications, like when the prompt, when the... You type in a prompt, Cloud Code runs. That's called the end of a turn. Okay, at the end of a turn, that's when your post hook stop notifications happen. For example, I have a post hook stop notification. When Cloud Code is finished running, it emits a mooing sound. So if you've been on my live stream where you heard like lots of mooing, that was like Cloud Code running and it was done. So what they do at slash goal is... On that stop post hook notification, they have another AI model. The default is Haiku. It serves as a small fast model to check whether the goal condition is met and decide like, should we keep going or should we stop because the goal condition has been met? So that was a technical explanation that you can also cut out of the YouTube video. But for those of you curious how slash goal actually works, it's like it's actually a wrapper around the stop post hook notification. And it runs that a small prompt at the end of each cloud code turn to check if the goal has been met. Okay, so here's how you type it. So let's just do a simple example. So how you type it is really simple. So you can just type slash goal. Okay, and it should turn blue. like this assuming you have cloud code in the terminal and it's updated make sure it's up to date okay and then you can see here it's trying to help you with documentation like what is the condition that you want to keep you want ai to keep going until this condition has been met so here's a simple example we're going to do sort every file in my downloads folder into subfolders by type keep going until no files are left do not delete anything and stop after 30 turns. I'm just gonna pull up my downloads folder somewhere. Where did it go? Where is my... oh no that's not my downloads folder. Oh here it is. Okay. Okay so so let's do let's just paste this. You can take a screenshot of this so let me just put it here so everybody can take a screenshot. Actually I'll type it into the chat. That's too long to type into TikTok chat. Okay. Yeah, no, sorry, TikTok. I'll type it into the YouTube chat. Sorry, sorry, TikTok. I'll type it into the Instagram chat as well. Okay, that... Okay, cool. So I typed it into the YouTube chat and the TikTok chats. Or not the TikTok chat, the Instagram chat. Now I forgot what I was saying. Okay, cool. So here, just trying to... I need to make it like readable. Okay, so here's my downloads folder. I'll put it down here. It's kind of messy. Now let's just run it and see what it does. Hang on. Let me move this here. Okay. So, oh, I forgot I actually already have a cleaner skill, but you won't have a cleaner skill. So when you run this, you know, you're not going to see like the goal matches my cleaner skill. But yeah, so now it's found 17 files. It counted how many are in each category. Now it's creating the folders and moved everything turn by turn. Okay. And pretty much done. Found 17. Let me create the folders and move everything. All files moved, zero remaining at the top level. Now it's going to verify its work, verifying the counts. Done. Your downloads folder is sorted. No files left at the top level. Nothing deleted. Boop, boop, boop. Okay. And then you can see goal achieved. Okay. So this one was a very simple example because I've gotten feedback that my examples are too complicated. So I'm trying to stick to simple examples. Okay. So it was actually completed very quickly. However, I have run goals and loops like this where it's running for five hours continuously. So don't let how fast this happened fool you. You can set quite an ambitious goal and let it keep running for hours on end. But here it was just a really simple example and now you can see my downloads folder has those other folders organized and everything was organized here. Videos, other images, etc. And so let's think about like just conceptually what happened here. So here are the things, here is how like that prompt was structured. So Claude was moving each file at a time, checking what's left and moves on to the next file and repeats. It stopped on its own when the entire downloads folder was sorted. And there were three things that made this like a good goal prompt. So let me scroll back up to it. Okay. So we have a clear end state. right like keep going until no files are left like that's a clear end state goal it had a check it can run right counts the loose files until still in the folder and a guardrail do not delete anything and stop after 30 turns loose i don't like this word loose okay i'm just deleting this from my love in cool okay so here's our goal sort everything keep going until blank this is the end condition so when you type this um the small fast model which is default haiku is going to check that no files are left like that's one of the goal conditions that it must check and then these are constraints like or guardrails don't delete anything and stop after 30 turns because with a malformed goal prompt. It could like keep going for hours and hours unintentionally. And so it's nice to put some kind of guardrail like stop after 100 turns just so you don't blow through all of your tokens for the month or for the week. Okay, cool. So that was a really simple example of a goal. Now I'm going to walk through some intermediate examples, although I don't actually have any of these receipts, but I'll just talk through them. Okay. So here's one. So here are five intermediate level examples. You don't have to run them right now, but I'm just giving you ideas of what else you can do. So again, standardizing a lot of messy files until every name matches a specific format. For example, I have a ton of invoices and receipts, okay? So you can rename them all in this particular format, in date order, and keep going until no file has its original name. Do not delete. or move anything and stop after 25 turns. So this is one example. The end state here, like what is the goal condition that our checker agent can verify? We have zero files left with the old style name. Now practical example number two, categorizing a spreadsheet until no row is missing a label. So for example, let's say you have a big spreadsheet and you're trying to label every row in the spreadsheet. So the goal is Fill in the category column for every row in expenses.csv. Here are the categories, food, travel, bills, shopping. Keep going until no rows blank. Do not change other columns and stop after 30 turns. Right, okay, so this is the goal. This is the end state. Keep going until no row is blank. And here are our constraints or guardrails. Don't change other columns and stop after 30 turns. Notice here, like we're not typing massive multi-paragraph prompts. Okay, that's kind of what people mean when they say like prompt engineering is dead. It's not really dead. It's just this is the next evolution of prompting. It's like how do we define... this loop so that an agent loop can like verify its own work, check what else needs to be done, prompt itself to figure out how to complete the work, and then check again if the work should be done. So the end state there was making sure there's no rows missing a label, zero blank cells in the category column. Here's practical example number three, working through a stack of documents until none are left. This pattern is called empty the queue. I'll talk about this in more detail when I share some cloud code routines I have for customer support, like processing every open ticket until none are left, or processing every email. or processing every unread email until none are left. That's like the empty the queue pattern. Okay, so here, slash goal, write a five-line summary of every PDF in my reports folder. Keep going until every PDF has a summary. Do not edit the PDFs and stop after 40 turns, okay? So again, we have a clear goal, we have a clear end condition, and we have our guardrails. And the end state is every PDF must have a matching summary. Practical example number four, cleaning up a bunch of social media captions until they fit all of your brand voice rules. For example, rewrite every caption in this file so that it meets this character limit and has no hashtags. Keep going until all of them are done. Don't touch any other file and stop after all of these turns. The end state is zero captions over 150 characters or containing a hashtag. And then this one is turning a list of rough ideas into hooks until the list is done. Turn each of these ideas into hooks, keep going, don't change other files, and stop after 30 turns. All 20 of these can be rewritten as hooks, less than 10 words. This is the end condition. Okay, so here's the cheat sheet number two. So this is just generally the framework to think about when writing your goal prompt. Slash goal, what done looks like. Checked by how, like how do you prove it? Is it a count? Is it that all of your tests pass? Is it that you have zero bugs in this area? And stop after X number of turns. Okay, so I know that these are, to me, they're simpler, okay? But again, I got a lot of feedback saying I need to have simpler examples. But this is the foundation for you writing way more complex autonomous agents. It's this framework. It's this cheat sheets of prompting. It's like not even about prompting. It's like how you are defining the task. Like, what does the end condition look like? How do you verify it? How do you prove that we've met it? And some kind of guardrail. Stop after X number of turns. Okay, so practice that. like try it with some of these examples you can tweak them so it makes sense for your use case or industry but like really get in the hang of thinking in this way you want to think in terms of loops where there is a verification step in the task itself because a lot of people actually don't ever add a verification step um it's something that i teach right like don't stop until but having observed thousands of people prompting i can tell you that the average prompt i see is like four words long so um if this is basic to you that's a good thing um don't worry we'll get into the advanced stuff okay so yeah hopefully everybody took a screenshot of this so like keep it as simple as possible because you're actually retraining how you're prompting. Okay, so like I don't care how simple it looks to you, just try it in this framework, keep trying it in this framework. Obviously, if you're a pro, you're going to create a skill that's going to help you write this prompt properly. But yeah, just keep doing it in this framework so that you train your mind to prompt in this way. Okay, so now we're going to build our first autonomous agent with a cloud code routine. So slash goal is great. It runs while you sit there. But a cloud code routine basically adds that timer. It adds the automation aspect that I talked about in the first concept of loop engineering so that the loop can run without you. Remember that the whole goal of loop engineering is to remove you, the human, from the loop. It's about how do we set up a system so that it can just keep looping until... a complex goal is achieved. So a cloud code routine is really awesome. If you come from an automation background, like you've used N8N or make.com or Zapier, I wouldn't say it's a total replacement for them because it's a little less deterministic than those workflow platforms, but it's... super awesome thing to have in your toolkit. So a cloud code routine is basically just a saved AI loop that runs by itself on a schedule, specifically in the cloud if you're using a cloud code routine, even when your laptop is closed. Okay, so you don't need your computer on all the time to run a cloud code routine. It's actually running in the cloud somewhere. It sets up a fresh environment, pulls your GitHub project, and then runs the skills. It can use MCP connections. It can make API calls, but it runs in that isolated environment. Quick aside, so if you are advanced and you like the concept of Cloud Code Routines, but you wish you could run it locally, just check out launchd. Tell Cloud to set up a launchd schedule for your tasks. And that way your routine on your computer runs and has access to all of your local files and setup. So if you don't want to run this in the cloud. Okay. Okay. So to set this up, we're going to go to clod.ai slash code slash routines. Okay. So that's over here. So let me just type the link in the chat. clod.ai slash code slash routines. Okay. Okay. Okay. So yours will probably look totally empty. I have a few here that we're going to walk through as examples. Okay. Oh yeah. We're going to do this one. Daily email cleanup. Okay. uh we'll do this simple example so but let me go to the edit so go ahead and click like new routine up here your screen is going to look like this i'm going to pull up uh when i already have and you're going to copy it and we're going to talk through what it does okay so click new routine and then you should have a form that looks like this again you can think of a routine as a loop that is running on its own in the cloud. You just set up a trigger which could be a daily or weekly schedule, it could be when a GitHub webhook fires, or it could be you can trigger it yourself sending a post request. We're going to keep it really simple and create a daily email cleanup agent that will read your unread emails and post the three most important ones to Slack. In my case, I'm just going to DM myself. one line each and here's the guardrail do not reply to anything okay so i'm going to paste these instructions uh in the chat oops okay you won't have sabrina obviously in your slack so you can just change that uh okay cool okay so that's the this is the name of the routine these are the instructions This is a very simple one. We're going to read all of your unread emails, figure out the three most important ones, and send a Slack message to myself. And do not reply with anything. This is the constraint or guardrail. This is where you can select a project, but you don't have to. So in this particular case, I don't actually need it tied to any particular project. It's just reading my emails. This is where you select the model, okay? So Opus 4.8, et cetera. And this is the trigger. When do you want this loop to run? Okay, in this case, we're going to say run daily at 9 a.m. So just click schedule and then run daily at 9 a.m. You can edit this as well. Okay. Oops. Okay. And then connectors here is what this loop will have access to. So remember, one of the key concepts of loop engineering is connectors, the ability for your AI agent to use the tools you use. For this example, we're just going to use Gmail and Slack to add connectors. You can click add connector here. And we will cover how to add tools that don't have these nice connectors and stuff like that. We'll cover that later. Okay. There are additional options here. We're not really going to cover behavior, notifications, permissions. Okay. So just make sure yours looks exactly like this. So name of the routine, simple instructions, daily at 9 a.m., and then these two connectors. And then go ahead and click save. Okay. If I make a skill, can that be triggered from a routine? Yes. So we were going to cover that. This is just a simple example. Yeah. So I'm just leaving this up here so people can make sure it looks like this. Okay. So, well... click save so then you'll have it here this is where you'll see um all of the routines you have running so i obviously just created this one right you can test it by clicking run now so go ahead and click run now in the top right corner and it will say workflow run started Okay, you can actually click this to expand it and see what it's doing. I'm going to close this because it's going to reveal a lot about my email inbox, but in a second we should get a summary in my Slack with my top three most urgent emails. Okay, I'm just going to check here if there's anything else I should cover. Oh, okay, we'll get to that next. Yeah, if you have questions and stuff, this would be a good time to post it in... youtube or tick tock uh okay this i don't know how long i do have a lot of unread messages in my email okay searching for gmail and slack now it's fetching a bunch of tools now i have the emails Now it's going to assess the most important emails like you can see its entire thought process here and by the way, you can continue the conversation So let's say it did something. Maybe you want more information You can come here in routines open up the chat and continue talking to it Like you can ask questions here. You can have it run additional things, etc. Okay Okay So if your connector is not listed, you will need to go back to the actual Claude.ai interface. So go to Claude.ai. So if you go to Customize, Connectors, and then this is where you can add your connector. So click the plus. Usually for most apps, they have an official connector, so Gmail here. Okay, and then once you add it, it'll look like this. A lot of people ask me about permissions, so this is also where you can manage the permissions of all of your connectors. So always allow needs approval or blocked. If you're a beginner, I recommend keeping it at always allow for read-only things, meaning like it can read your content, but it needs approval if it's going to write, edit, or delete anything. Okay, so that's a pretty good configuration, especially if you're a beginner. Okay, let's go back to our routines and see if that finished running. Okay, so that one finished running. So you can see checkmark here today at 1244 p.m. So let's go to my Slack and see if it DM'd me. Yes, so. Okay, so here are my top three unread emails. You can see the timestamp is literally right now. Someone sends a long book excerpt. Okay. Which plans for Blotato showcase at some event? Okay, interesting. So these are my top three most urgent emails. It was sent to my Slack. Okay. And again, you can open the conversation here. You can continue the conversation if you want to update anything. And you can update the... the routine itself. So once you create it, it's kind of going to look like this. Click this pencil button in the top right corner if you want to go ahead and update it again. Okay, that was the simple example. I'm just going to talk through intermediate examples and then I'm going to show you an advanced example, which is like my own routine. Okay. So here's some intermediate level examples. Let's say you do have like thousands of unread emails. It can completely sort your inbox, label every single email and archive the ones you don't want. Okay. So you can set that up as a routine running on a daily or a weekly basis. Honestly, this one's really awesome. uh just just to help label emails even like label receipts label filter these out, archive these emails. So if you have a monthly report, you can create a cloud code routine that every single month pulls data from certain systems. Let's say your CRM and your email marketing system, and it'll write the first draft of your report and send it to you in Slack. Okay, so that's another idea of an intermediate level example. If you have a lot of news you're trying to keep up with, you could say every Monday, read the updates on this topic and hand me a short brief so I stop. doom scrolling and but I stay up to date on the news that I care about you can have another cloud code routine that does this so what I would recommend is picking one of these or a variation that you resonate with so you can make your first routine and make sure that it's running autonomously every day or every week okay so now let's walk through a real example because I know people have questions like well how do you integrate skills what about other apis etc so that will all be answered here so Let me just pull it up. So now I'm going to show you a real example of a cloud code loop, including the main agent that does the work and the checker agent that independently reviews the work. So it's going to be these two. I wish I could move them next to each other. So this one, every day, runs at 8 a.m. and cleans up support tickets and also surfaces any issues and gives product roadmap recommendations. this is the checker so it will review all of the closed tickets that were closed by this agent and confirm they should have been closed because i don't want to close support tickets that should have been escalated to a human okay so let me open this one up click in the plop okay cool so here's what this looks like so it's actually it's actually not that much longer but it utilizes different things. So here, I have a skill already in this GitHub project. By the way, this is where you can add your GitHub projects. For me, help.blowtato.com is where I host my support documentation, and it also has my skills related to... customer support. Okay, so run the cleanup ticket skill on all open customer support tickets using the Intercom API, which I'll talk about as well. Provide a summary in Slack, hashtag support channel. And when referencing conversations, always provide the full link so I can easily click and open it do not delete tickets here are constraints and guardrails and do not stop until you've processed all currently open tickets this actually isn't necessary because this is already the goal condition but just one it's like wanted to just make it extra clear here here's the trigger this runs daily at 8 a.m And then here are my connectors, intercom and slack. Okay, in this routine, I also actually have an API key. So what you can do is you can create a separate environment that stores like your environment variables. So the default environment for me does not have any environment variables. And then my support environment contains my API key for intercom. And the reason for that is using the intercom API, access to a lot more things than using the intercom mcp for example the intercom mcp does not allow me to close open tickets but the intercom api does so for those of you who are feeling limited by connectors or you want to integrate an api that doesn't have a connector for example i'm i don't think perplexity has a connector to clod since they're competitors maybe now they do but you can you can just create a new environment like click add environment and then you can put your environment variables here so let's say this is our research environment And we want to put our perplexity API token here equals and then put the token here. So this is what it would look like. So any tool that does not have an MCP that fulfills your needs, you can just integrate the API into the routine just like this. Just create a new cloud environment, call it whatever its purpose is, and then put your environment variables, your API keys and tokens right here. So I have one for support. And then that's it. So it will run daily at 8 a.m. It will analyze all of my open customer support tickets. It will automatically close the ones that no longer need help or need to be escalated. It will also look for any urgent issues like... outages, bugs that are urgent that I need to fix. And it's all tied to this cleanup ticket skill. Okay, so there's actually a lot in this skill that I'm not showing. But that's the whole point. Like you can create these skills in your GitHub project, and then your AI agent loop just can use these skills in the process. Okay, cool. So, okay. Let me close this. So this is the maker agent, the primary one doing the work, classifying tickets, analyzing them. And then I have the checker agent here. So this will run about two hours after the other one. And its job is to independently verify all of the support tickets that were automatically closed today. Confirm that each ticket should remain closed. If the support responses received by a customer do not clearly solve their problem, reopen the ticket. So again, in loop engineering, the concept here called it sub-agents. One primary agent to do the work and another agent to verify the quality and accuracy of the work. So this is the checker agent that helps verify it. So it's set up similarly, right? So the only difference really is it runs two hours after the first one runs because the first one takes a while to actually process all of those tickets. And then you can see examples of it in Slack. So if I go to support, So, okay. So here's the ticket cleanup run. Okay, just a lot of tickets. Okay. And then let me open this. So this is the cleanup checker, and it was running. So, okay. So here, I'm just pulling up a summary. Okay, yeah. So here in the first run today at 10 a.m. that I triggered manually, the full audit was complete. 151 tickets were currently closed. Or 151 tickets were correctly closed. One ticket needed reopening. Okay, so this is an example, again, of the checker verifying the work, and it found that one ticket needed reopening. Okay. So that is a more advanced and realistic implementation of an autonomous agent. So I'm just scrolling here. Yeah, and then, I mean, I covered this in the blog post, but the next level is to integrate any API with your routine, which I talked about a little bit here, right? Like if you want to edit this, click this cloud, and this is the environment. Click add environment, and this is where you can put all of your API keys for any tool you want your cloud code routine to access. Okay. And then, yeah, just some tips here. Like if you are brand new to all of this, just start read only. Like instead of having it reply and close tickets, just have it summarize your emails, like label your emails. And that's why I started with that as a simple example. Like I know it's this for people in AI who've been doing this while, it's like. a really simple example however again i've gotten feedback that my examples are too complicated i should start with simpler examples so that's a nice one because it's just reading your emails it's not replying to your emails it's just reading them labeling them and surfacing urgent emails and make sure you set your constraints and guardrails you know do not reply to email or do not delete anything in my case do not do not delete any support tickets And then routines use the same plan as your normal Claude, and there's a daily limit in settings, so just be mindful of that. And that's another reason why we have a guardrail like, you know, stop after X iterations. So here's the cheat sheet number three. So ship one tiny routine this week. Think about when should it run? What is the trigger? What should it do? Fill that in here. What are the tools it needs? Connectors and API. Make you can make a list here and what is the limit that you want it to have like don't reply to emails or don't delete something okay, so Like think about this cheat sheet and fill it out and this will help you create the clod code routine By the way Claude can create the routine for you if you like literally copy paste this into clod code and be like create a clod code routine and we fill out these blanks, it will create it mostly for you. I think you will need to go in here and still create your environment variables and stuff. But OK, so yeah, start read only. Let it summarize stuff for a few days before you let it write anything, edit anything, or delete anything. But definitely use this template to, again, reorient how you've been prompting so that now your prompting is all about creating these loops. Okay, so just to recap everything we've talked about. So we talked about loop engineering is a shift from typing prompts to designing this loop. Slash goal and routines are things you can use to make loop engineering a reality much more easily in 2026. So slash goal, and then you type the goal you want it to keep going until it hits that goal. And then there's going to be a second verifier agent that checks whether that condition is met. And then a routine, it's like a timer. And that can be connected to your skills, your connectors, your API keys, your notebook, which is your GitHub repo. It's connected to all of those things, which are all key concepts in loop engineering, so that it can run the loop in the cloud without you. And that will be your first fully autonomous agent where you are removed from the loop. Now, the hard part, I really try to emphasize this. The hard part is defining... what done is like what is good what is good enough what is done um that's why like my personal formula whenever i think about how am i using ai effectively the amount of ai leverage i personally get is always a function of like my skill and my clarity this is like the formula i i personally think about all at the time by clarity i mean my ability to define what it is I want it to do and what done looks like. What does good look like? And skill is like my ability to review the AI agent's work so that I can further improve the loop. So a lot of people like this simple formula because it's cool, easy to understand. But this really is how I think about AI. Like if I have no skill in something, you know, AI can help me build up some initial skill fast, but I'm just not going to get the same amount of leverage as someone who is already very skilled in the thing. This is why like senior developers are like so much more productive using AI tools than a junior developer. Because you have the ability to review the work, identify clearly like what is wrong and how the loop should be improved. Like maybe you didn't clearly define what good looks like or what done looks like. Maybe your end condition was make sure all the tests pass. but your tests were garbage, right? Like you have to go improve the tests and that requires skill to be able to like recognize that and do it. So this is really my personal formula whenever I think about how do I use AI to maximize my leverage? Where do I have skill? Where do I have clarity? Like to be able to define what it is I want and what good and done looks like. Okay, so yeah, that's pretty much it. So hopefully that was helpful. I guess I'll just do another recap. here with all the lines where's my loop okay man this video is so cool it was made with ai um i'm looking for the graphic with a circle yay okay cool so here we are um okay cool so yeah so okay let me so So in the beginning, we talked about like, what are the key concepts of loop engineering? Because like, that's the latest buzzword. The reality is people have been doing loop engineering, we just didn't have like a collective name we all agreed on. For many of you who have been following my content, you should be familiar with at least half of these, right? So concept number one in loop engineering, automation, like running it on a trigger or when something happens, do this. That way, you're not the one sitting there prompting it to happen. It's just happening without you starting the loop. Number two is work trees, which means let's say agents are working on the same code base. You want each one to have its own safe space so it can experiment without colliding with all of your other agents. Get work trees. Practically speaking, like most people don't have to do anything to set this up. And practically speaking, most people won't. really need this unless you're really having a lot of agents modifying like the same exact things. Number three is skills. So many of you are familiar with this. It's basically a playbook that your agents can follow to do a task the same way every time. And so you don't have to give it that context and instruction every single time it does a task. And number four, connectors. Just like we saw connecting it to Gmail, connecting it to your support desk, connecting it to your CRM, connecting it to your content creation tools, connecting it to Plotato, for example, to publish the social media. Connectors are what give... your AI agents the ability to use useful productivity tools, the ones you use every single day. Number five, sub-agents. The key concept here for loop engineering is you want one agent to do the primary work. It's usually the smarter, bigger model. So let's say Opus Sonnet 4.8 or Opus 4.8. And then you have another agent that's typically the smaller, faster model, like Haiku, reviewing the work independently. Has this goal condition been met? Yes or no? Haiku is going to check that. If the answer is no, the primary agent has to go and keep trying until it completes the goal. And the last concept for loop engineering is memory. It's like a shared notebook for every agent run to have the same context of what's been done, what's been tried, what are we struggling with, what needs to be done next. Practically speaking, if you have a GitHub... project already hooked up to your Claude, that is going to serve as the memory as the scratchpad. Okay, so those were all six concepts for loop engineering. And then we also talked about how to use Claude code slash goal to make your first loop and also how to use Claude code routines to make your loops autonomous running without you. That's it. Yeah. Oh, man. My voice is... gone but okay if you have any question i'm gonna stop recording now uh but if you have oh wait obs is on my other screen okay like i'm gonna stop recording otherwise the file is gonna be really large but yeah if you have any questions just drop them in the chat um i'm checking both youtube and tick tock chat right now and i just have to go stop the recording okay sweet oh oh okay so yeah drop your questions in the chat while and let's just like compile the questions. I need to order sushi real quick because I'm really hungry. But I'm also on a diet. Okay. So, okay. But yeah, you can drop your questions in the chat and I'll check in a second. Oh, wait, let me ask if I need to ask if my friend wants sushi too. Where is the template? I am going to drop that later today. Just make sure you're subscribed to my newsletter, which is my link in bio. Yeah, the cheat sheet will definitely help. Just built a playbook. What's the difference between doing this and in cowork? I think cowork doesn't have the slash goal command yet. So as far as I know, that's just in cloud code. But you can definitely do pieces of it in cowork, right? So cloud cowork has the schedule thing that is similar to cloud code routines. So you can schedule recurring tasks and stuff. What is the best way to start from the beginning to auto post with Claude through Canva? What is the best video? Oh. Let's see. I think I have a Canva video somewhere. Yeah, probably this one. Okay, I'll drop this in the chat. This live stream, it will probably take a week for the editors to edit it, unfortunately. Is there any AI that is actually free? Yes, open source models are free. You can try setting it one up with like LM studio as one kind of app. Okay. Can you stop a goal in the middle of the execution? Yes. Just type slash goal clear. Okay, like this. And also if you type slash goal while it's running, it will give you a status update. Like it'll say, here's what I've tried and here's what we're trying to do, et cetera, et cetera. So you can actually type slash goal while it's running. Okay, my friend's not replying to me about sushi. Should I just order her sushi? Just in case? Okay, yeah, I guess that's what I should do. Okay. Oh yeah, I need to get my eel unagi. Oh, I'm so hungry. I stayed up till 1 last night because this Hyperframes was so fun editing this video. Okay, cool. What was the skill you used for the presentation? I'm not sure what you mean, actually. So all of the prompts will be here in this newsletter, which I will send out literally like today after this live stream. I just want to check it one more time. Am I charging 10k for this? No, it's actually free, but it's worth 10k. So I put 10k in the TikTok title. Which totally free will create animations at least three minutes long? Check out hyper frames, remotion, and LTX. Oh, hang on. Where is the questions? If you have questions, now's the time to drop them. I have 20 minutes left. Oh, I'm a video editor. Oh, yeah, that's okay. I have too many video editors now. Why do I do TikTok instead of consulting for companies that will pay you loads of money? That's a good question. Honestly, it's not like that fun. Like, I've been at a big company, and honestly, all of the productivity gains and improvements that you deliver don't trickle down to employees. Like, they really go to the shareholders and the people at the top. So, I am more interested in, like, just... democratizing this knowledge on TikTok for like normal people so that you can make your own money because I really don't believe it will trickle down to you even if I like yeah I can teach a company it will save them or make them billions of dollars but like that's not gonna help you the employee so if you think so well that's unfortunate for you but yeah yeah I mean I also yeah I don't know a lot of people ask me if I want to do AI consulting but it's first of all i'm very introverted i don't like interacting with people so like yeah i'd rather just sit here in my shitty corner so um how many people do you think will use these loops i think it's it's really not common mainstream knowledge at the moment like even before cloud code goal and routines like You know, I frequently tried to prompt don't stop until this, but it's, it's not perfect either because it's using the same agent to do the grading, which as we, as we see, you know, it, it, um, it just lies and hallucinates that it did a good job. Um, so not many people are like really using loops in this way, like 99.999% of people I see, I swear. It's just, they type a prompt, chat answers. They read the answer. Then they think about what they should prompt next. They type the prompt. They wait for chat to finish again. They read the answer. Then they think, oh, what should I type next? Like this is the way 99.999% of people use AI today. Like for sure. If you don't think that, you probably exist in a bubble and like should go talk to like real people. But yeah, real people, that's how they are prompting today. Very few people are really setting up. loops like this that are running autonomously almost all of the time. And it's not that it's like autonomous forever. Like you do have to be very involved in reviewing the output and improving the loop. So even for my customer support agent, I'm very involved in like figuring out, oh, it got this wrong, it got this wrong. Or like I just changed the product and some parts of my help docs or earlier YouTube videos are now out of date. You know, it's like in a business context, like it's a very dynamic. problem. Like I'm always improving systems, ripping about, ripping out things that no longer are working, replacing them with a more efficient way of doing things. Like it's changing all of the time. Um, okay. So you have a way to use AI for content generation. Yeah. That's what I was showing earlier in this thing. Hang on. Where did the video go? So yeah, that video I was showing was like 100% edited with Remotion. So, draft. I'm working on storyboarding it as well. So like by storyboarding, I mean, so basically like I'm getting a lot of YouTube. Wait, can you guys still see my screen? Yeah, okay. So basically by storyboarding, I mean, I'm getting a lot of feedback that I should spend more time. pre-production like scripting my videos etc so that they're clearer and so I've actually been experimenting with using this free AI tool by free I mean like it's 100% free the only cost is like Claude to call it but you can use an open source model instead of Claude and so I'm experimenting with a pipeline that storyboards a video and then I film my talking head so that way I can like see like do my words are my words landing and is the visual like helping to clarify what i'm saying so this is an example storyboard actually so i fed it the same script but i didn't give it my video so here it's actually just storyboarding like what each scene should be like here this is going to be my talking head okay and then this other scene is going to be this visual and then it actually uses um text to speech in hyper frames so there's actually voice so i can hear the the AI voice, not my voice, but I can hear the AI voice with the visual and then make a judgment like, is this landing? Is this the simplest way I could explain things, et cetera. And so this has been really interesting actually to experiment with. Again, everything you're seeing on the screen is 100% free and AI generated. And you can hook it up to an open source model so you don't even have to pay for cloud tokens, right? So So yeah, this has been a very helpful exercise for me because I'm not a visual person. Like I enjoy writing. So I actually don't mind spending more time like writing and prep. But it's very hard for me to translate like what I'm saying into visuals that. actually clarify the point. That's why I usually just have lots of text open on my screen because that's how I process information. I literally don't watch YouTube videos. I would rather read help docs than watch YouTube videos. That's just how I learn. And so it's something I struggle with for YouTube. How do I convey some of these complex concepts in visuals that are easier for visual people to follow? That's why they're watching YouTube videos. Yeah, I know. A lot of people are the opposite. A lot of people get overwhelmed when they see a lot of text. Like, I am so weird. Like, I read every single legal document ever in my first startup. Like, I read it multiple times. Like, and I enjoyed it. Like, that's weird. Like, I know that's weird. Like, I'm super into reading and writing, and I'm completely not visual at all. Like, if, like, an editor is like, oh, can you give me feedback on this video? And I'm like, well... I don't really know what to say. Like, I guess I like this part. I got bored at this part. Like, I don't have much more feedback beyond that. Um, so this working on this has just like helped me personally like bridge that gap between like, what am I saying in my script? And is the visual helping to clarify what I'm saying? And I just started this like last night i was up until like 1am working on this um i'm calling it my storyboarding pipeline storyboard skill right so it's it it takes a script and then it creates like the visuals that are going to be with the script and then i'm going to go back and tweak the script if i'm like this wording isn't landing with a visual or this visual isn't landing with the wording i need to change one of them or both of them Yeah, this is just called HyperFrame. So it's super easy to set up. So if you just tell Claude, set up HyperFrames and create a two-minute horizontal YouTube video just about loop engineering, whatever. Like, just put that one sentence in, and in 10 minutes, Claude code will be done, and you will have a, whatever, two-minute YouTube video about loop engineering. So yeah, just do that. Okay. Best... Full pipeline. Is PC or Mac better? Honestly, if you're non-technical, I generally recommend Mac. If you're really technical and you like playing around with open source models especially, you will probably get a PC with Linux, right? Like some kind of Ubuntu flavor. Yeah, this is Hyperframe. So for a while I use Remotion and I have a shorts editing pipeline in Remotion, but I was playing around with Hyperframes last night. So the difference is Hyperframes is HTML based and Remotion is not. I believe it's React based. So you can extend Remotion like very significantly. Like you can have Remotion make 3D stuff with 3JS and stuff. You can have it make like cool math visualizations with NM. Like you can extend it. really significantly whereas hyper frames you can't extend it that significantly but it is easier to use out of the box so if i had to make a comparison it would be like hyper frames is apple or iphone and remotion is android google okay so choose your own adventure there but yeah that's based on my um testing so far um it's really fun though it's it's teaching me a lot about Like maybe I should have more visuals in YouTube videos to like clarify complex concepts or not. I don't know. Like I'm just I'm trying all this out. In my ideal world, I would just stand at a whiteboard and just like talk all day about AI. But, you know, it's helpful sometimes when there's like actual visuals that are not like my hand-drawn chicken scribble. Um, how can I become your YouTube editor? Honestly, I have too many YouTube editors and I'm slowing down on video production because, um, well, basically I'm just getting advice that I should spend a lot more time scripting and pre-production. So for the next month, that's what I'm going to try. Yeah. So like, honestly, if you just want it to like, look really good like this with as minimal work as possible, use hyperframes. If you, um, are like... If you really want full control, you're going to have your own design system for your video pipeline, you want to extend it with other tools like 3JS for 3D visualization, then invest in ReMotion. Far more extensible, okay? But I'm guessing most people have an iPhone for a reason. Because it's like, we don't need all that. We just need to make a decent video, okay? So we love my hand drawn info. The thing is like even when I vibe board, so if you have been to several live streams, even when I vibe board, it's like I just I just write text. I'm like I'm literally so not a visual person that when I whiteboard, I just write text, you know? So but it's been fun. It's like it feels like a. stretch goal. Like, oh, I have to really figure out what, like, what is the best visual that will convey this point? You know, like that's, it's hard for me, but it's use, I think it's useful to learn that skill. So that's why I'm investing in practicing that. And AI is helping me with it. Like, again, this storyboard was completely generated by AI with a voiceover. So I can like literally hear like, oh, that. doesn't quite land or the visual is has the wrong timing like the visual should be later a little bit um or it's kind of confusing here like why am i saying these things but the but uh the text says these things like you know there needs to be like coherency uh coherence in the thing so um okay yeah a lot of people switch the android for the camera because it does have a better camera But then they're like, screw it. I miss Apple. Like, everything just works in Apple. So, I am doing an in-person meetup in the school. We literally have an in-person meetup next week. So, if you're not in the school group already, like, I'll post the link here on YouTube. But it's literally the best free community you will ever join, I promise. And if that is ever not true, I hope somebody yells at me. How has full stream marketing operations gone with this for SEO, social media? Yeah. So like how my agents, okay. So how my agents actually look on a day-to-day basis is I have like a terminal open with three to four agents, each one working on a goal. So like. I'll have it be running throughout the day. Then I'll remote control so I can just like walk around the house and do other stuff. So it's like literally just running, working on the goal. When it's ready for me to check something, I check on my phone to see what's being done. If it needs a lot of feedback, then I'll go to my computer and be like, okay, we need to fix this, fix this, like how I set up the goal. had incomplete information. But I literally have a go-to-market engineer, a contents engineer, a YouTube engineer, and then a support engineer pretty much working daily with their own respective goals. So in terms of full-stream marketing, I mean, it's still early, right? Like, you know, slash goal is brand new, cloud code routines are brand new. It's still early, but it's just like, to me, I see like a massive... productivity gain. Like while I was making this thing, I had another agent write my whole newsletter with the instructions and we went back and forth. Like, how do we simplify this further? How do we, let's make the examples even simpler and build up to like more complex examples. Like, you know, it still takes feedback to do it well, uh, in terms of having multiple streams for marketing, but it's, it's like incredible how much tedious stuff. it can automate um and i'm just i'm really there to provide like the high level this is our strategy this is where we want to get to and provide feedback along the way like okay you were i clearly didn't specify enough information here and i need to go back and add context so that the ai agent loop has enough information to know when the job is complete so um it's not mill it's not millions of dollars um i don't know what that meant the It's not millions of dollars, not even close. It's both. She runs Blotato. I was her admin? Who is... I'm confused who is my admin. Okay. Well, okay. Hella set up an in-person two-days class helping each person set it all up. Well, that was the point of today. You guys are supposed to follow along today to set it up. So... What's a voice linter? You mean like for if you have an AI voice? Is it possible to productize and package something to sell to SMB? Yeah, of course. That's literally what my product is. I analyze frame by frame my stock footage and edit. Oh, well, FFMPEG is very limited for editing. Like FFMPEG is a good utility library. And it's what a lot of these tools use, by the way. They use FFMPEG. But you can actually tell Claude, like, analyze the video. You can also plug it into Gemini to analyze the video. Or you can tell it, like, take a screenshot at each three-second interval in the video to analyze what's happening in the video, you know? It's hard to see it going on to... Oh, yeah, I'm not really sharing anything right now. So we're just wrapping up the call until my sushi arrives. Oh, yeah, it's arriving in 15 minutes. Okay. So to Rachel, I have a video dropping soon that's like 12 ways to make money with AI. And I really just tried to consolidate like... multiple different revenue streams and which ones are scalable versus faster income. I think it's going to drop this week. I hope so. But I would look out for that video and I will probably send a newsletter along with it. But like that video is the one where I really try to break down like here's what you can do, etc. Okay. I think I am going to just send this newsletter now. I mean, I don't I feel like these intermediate level examples are a little easier as well, but that's fine. Okay. I'm just reviewing this. I think this is ready to send. Okay. So yeah, if you are looking for the companion... Oh, I should have put a picture. If you're looking for the companion newsletter that has all the prompts and everything, I'm sending it out right now. So it should be in your inbox in a bit. So just look out for that. Someone is spamming to become my YouTube editor. Okay, that's kind of annoying. To sign up for my newsletter, just go to the link in my bio. So I just go to Sabrina.dev. It's not that I don't want to hire new YouTube editors. It's just I already have YouTube editors. So you would have to be significantly better than my current YouTube editors, whom I enjoy working with. You know what I mean? Besides, I'm really trying to focus a lot more on pre-production for the next... couple of months and just like see how that goes so okay um so what are you analyzing each frame for i guess can you also analyze a photograph a technical drawing from an architect yeah you definitely can so um just just drop it into cloud it we'll analyze it uh well my school community is for women so who are building an AI and you do need a valid LinkedIn profile. This is how we confirm your identity. Okay, so you can apply to join. It's free. It has 4,500 women in there. We have educational trainings and workshops almost every single day of the month. And then we have an in-person retreat literally next week in Salt Lake, where I'm going to take everybody hiking. Any tips on social media listening skills for agents? How do you write your scripts? Yeah, so for that, I use Apeify and then just tell it to like, so Apeify, find the top 10 videos. on this topic you can put in keywords or hashtags from the past month with the highest number of views and the lowest number of followers so i have an air table tracking this i call it like momentum score because like sometimes the creator will just like have a lot of momentum on their on their growth and i want to see like their entire profile so um that's what i do so apify i'll just type it in tick tock apify fine top five videos on this topic or hashtag in past 30 days. And you can have that autonomously running, right? Like every single week, have that running, and then it will just email you or send you a Slack message with who those creators are and what those videos are. Okay. Any obsolete MacBooks you'd recommend to run locally or to do all of this you taught us today? Yeah, like anything you can afford, honestly. There was nothing I showed today that was locally intensive because the cloud code routines run in the cloud. Unless you're doing like launch D to run route. Oh, wait, hang on, my dog rumor. Yes, is texting me. Okay, yeah, I mean, you can run routines locally using like LaunchD if you're on Mac, but I didn't cover any of that today. So nothing I showed today is like crazy computationally expensive. So you can definitely just buy a used MacBook. In fact, when I buy a MacBook, I usually just buy refurbished because like I have so many already, like it's fine. It does not need to be brand new. Yeah, like I use my 2016 MacBook until like 2022. So, yeah, you know, it's generally not like, like, you're probably not running into a hardware issue unless you're doing what I'm doing, which is like OBS, four agents running in the background, streaming on five platforms. Okay. Yeah, I mean this is the women's community. I do plan to start like an open community not just for women at some point. It's just the problem is spam. I found that at least in the women's community like so many women are helpful in cleaning up. like spam and promotions and stuff and so it's it's been a like a tremendous crowdsourced effort to keep it high quality and um i just worry about that with a with an open community that anyone can join you know what i mean because like even for our community we have screening questions you have to have a valid linkedin profile that's how we ensure like people are real like not just like deep faking somebody else Where are the details on Salt Lake? They're in the school community. You can ask. You can create a new post in the school community because it's literally happening next week, June 25th to 28th. I personally prefer Pro if you're choosing between MacBook Air versus Pro. But it's like whatever you can afford right now. I think Air is a little too weak, especially because I do video stuff. So if you can get... pro then get a pro get a used pro um what do you do when your agents are running overnight and are well over 30 percent in their context window So if you just don't want them to burn through everything, that's where you can set guardrails, like stop after X number of turns. And you could say stop after X number of turns and produce a summary of what you've done, right? That way it just doesn't burn through your context window if you need the context window for the morning for your actual work. Are you backing up skills or anything to a hard drive? All of my skills are backed up to GitHub, right? So like every project I work on has a... has a GitHub repo. So like coming from a technical background, it's like if you have, you know, different code bases, you would have a different GitHub project. So I do the same. So like I have one for Blotato marketing, one for my content empire. I actually made a new one just for YouTube because I'm experimenting so much with like long form editing and it's pretty different from short form editing. So I have a new project for that. I have another one that's just my website and SEO. So blotato.com and it has like the blogs and stuff there. And then I have other projects obviously for technical things, playgrounds, my blotato app, my help docs. So I have separate GitHub projects for each like. area of my company and then each of those GitHub projects is synced to GitHub on the cloud. Each GitHub project has the cloud skills in it. So I back up the cloud skills in the cloud. So that way I can just like change the cloud skill. If I end up not liking it or want to change it back, I can easily do that or I can see what it was before and compare it. So not spamming, but really want to know how you think consultants should approach cold outreach? Well, always with value first. Actually, there's a really good story of a woman in our AI community who just shared that she basically like cold DM somebody on LinkedIn. and ran a security check on their website and found some critical stuff, like important stuff, sent it to the person who was cold. And that person was like, oh, like we need you as a consult. Like we need you to train our team on AI literacy and do this other stuff. That is cold outreach, but it's like leading with a lot of value. Like it wasn't just like, I DM you, buy this. It was like, you know, just DM. Hey, I see we're like in the same industry or whatever she said and then she happened to run the person's website through a security scanner found a bunch of vulnerabilities shared it with them and they were super positive in their response That's an example of like cold outreach that leads with a ton of value If you I blocked the guy who's a video editor But if you were a video editor trying to cold DM me the way not to do it is to spam my lives with zero value I would be really more impressed if you edited this live in real time. Like you recorded 10 minutes of this live and sent it to me as a link in the chat right now. Like that would be impressive. And I'd be like, oh. That's kind of cool. He just like did that live right now. Like that would be interesting. That would be leading with value. Even if I say no, at least I have like a good impression of you. I can refer you to other people because you actually took the initiative to do that. Lazy DM, lazy cold outreach is literally like this. Like you hop on my live and you just spam 30,000 times like the same message that adds zero value, you know? So just don't do that. So she actually vibe-coded her own tool to run these security checks. So she ran this prospects website through her own tool and then DMed them the results. Yeah, I mean, you'd be surprised. I read most cold DMs. on certain platforms like my instagram and facebook are insane because of my dm automations so i don't read those tick tock i don't read because it's just a really messy inbox and i get a billion but linkedin i don't have that many followers on linkedin so i still read the dms there um i still read my email inbox you know so people cold outreach me all the time it just it's just like lead with something valuable like what like you know so I currently use ManyChat, but I want to replace it. So I'm actually building ManyChat, but with MCP API so that I can just have Claude create my ManyChat automations. And that will actually be part of Blotato, hopefully by the end of next month. Yeah, there's definitely a lot of scammers, guys. you know just you like you literally cannot pay me um unless you sign up for my very cheap low ticket product which by the way you can automatically get a refund for if you don't like it just hit the support bot click refund you know like don't pay me any other way there's just don't send me investor money don't send me crypto like what just i need to make more announcements about that because people always fall for it um is it worth it considering older models Uh, yeah, if you're going to be running your own models, sure. Yeah, you would want, I mean, to be honest, if you're going to be running your own models, get like a Mac Studio. That's what I have, um, especially for video related stuff. And then I do want to experiment a lot more with open source. So if you can afford it, that's what I would recommend. But, uh, someone wrote to me as you yesterday on Facebook to join your group. Yeah, there's a lot of scammers on Facebook. Video, you know, okay. um oh hang on my my sushi is almost here uh if you i just put the link in the chat so it's school.com i'll just go there since i'm sharing my screen actually i'm not signed in here i forgot so so here it is it looks like this i should probably change this logo soon put put like i don't know a picture of me or a picture of us there something um but yeah that's the link i just dropped it in tick tock You're teaching Claude, but do you favor other services or think them superior? I generally prefer to just go deep on one ecosystem so that I'm productive in it. So I personally have been using Claude code for well over a year. Like I personally publicly announced I was permanently leaving cursor in favor of Claude code. in February last year. And by the way, people, I got pushback on that. People were like, no way, no way, no way it'll be permanent. You know, it probably won't be permanent. There will be some other tool that comes out that's better. But like, as of 2026, I find the, it's the ecosystem and integrations of cloud code that make it. Just my tool of choice. It's like a Swiss army knife for anything I might want to do. Like certainly I use specialized tools for other things, right? Like for my support stack, I use intercom as like the front end support widget also for my email marketing. And then I have local open source N8N literally hosted in the closet over here that runs my entire AI support bot for like $1 a month, right? Like that, like... That's just how it's set up. But like Claude is kind of my Swiss army knife. Like whatever I might want to try first, I'll probably try it in Claude first. Where it sucks is anything visual related. Like I know they just dropped artifact support in Claude code, but... it's still like really basic. So if you're doing a lot of image and video generation, for example, in Cloud Code, it's like hard to visualize. If you're stitching a bunch of video clips together, you have to like open it every single time and it's like kind of annoying. And only recently did Higgsfield come out with an MCP. You can use Replicate and Fali AI APIs as well, plug them into Cloud Code to generate stuff. And then Cloud can use FFmpeg to stitch everything together. But it's still like not an amazing experience. So anything... like visual that requires a lot of visual review is still not a great experience in Claude. So, okay. Yeah. Unfortunately that particular group is just for women. So, because I don't know, like the spam problem is really tough. I don't want to start. a community, open it up to everyone, and just like lots of spammers join. If you actually look at most free school communities, it's just like total spam. Like it's like, it's like, it's really just an email list at this point. So, you know, like no one's actually going into the school and like engaging because it's just way too much spam. So if you do not have enough space to run things locally, what cloud do you recommend for agents? So in the... Third part of this tutorial, we talked about cloud code routines over here. So that's over here. Like, OK, I'm not going to rehash it. Oh, I was wondering what the poll was so far. OK, how good. So this is a poll, by the way, at the bottom of my newsletter. So I'm starting to add these to see, like, is this too hard to follow? So that's interesting. 21%. This is interesting, but not useful. So it would be nice to know like why, but you know, Substack sucks. So like Substack has like zero automation or like workflow features. So compared to Beehive or ConvertKit, if you're coming from that. So, okay. I think my sushi is, okay. A few more minutes and then I'll hop off. Let's see. I'll go back to YouTube. Yeah, the newsletter is already sent, so it should be in your inbox. So make sure you check the spam folder in case it's landing there. I'll drop the link as well in the chat. Yeah, so Canva significantly improved their MCP connection with Claude. So I dropped some new tutorials on it. There are tips and tricks to make it better. Like, for example, if you're using a Canva template, I have Claude first look at the template to check how many words should fit into each, like, text block. Because if you don't do that, then, like, Claude will just... put too much text or too little text and it ruins the aesthetic of the template. So like that's one little hack to like make it look better. So I think the Canva MCP is actually quite usable today. But you have to kind of like set it up in that way. And so I do have like two or three Canva tutorials on my YouTube, just walking through like tips and tricks to get the Canva MCP set up properly. Because a lot of people will try the Canva MCP, and they'll be like, make this, and then it will suck, like for sure it'll suck. So the real way to use it is to use Canva templates and have Claude populate information into those templates. And then it's going to look way better, like way, way, way better. Okay. Oh, yeah. I posted the newsletter link here. I pasted it in Substack TikTok, and I'll paste it in Instagram as well. I personally don't use any CRM systems anymore just because I have a low-ticket SaaS product. I initially set up GHL for Blotato, but it was like... really unnecessary. Also, Intercom has basic email marketing features. So I can technically send email campaigns, I can set up goals hooked up to Stripe to see like, oh, in this email campaign, 50 people converted into paying customers, right? So it has like just enough email marketing tied to revenue that I don't feel I need to set up a separate system like Salesforce, GHL, or MailChimp. For my newsletter, I'm just using Substack because it's free. It's like a newsletter plus social network combined. So that's what I use. But okay, if you're not getting... So it takes a while to send it through email because I have like 230,000 people on my email list, but I just dropped the link here on Instagram as well. I will drop it again on YouTube and I will drop it again in TikTok. Yeah. Can you offer me scholarship under your AI school? Well, the good news is all my AI education is free, so you don't need a scholarship. You could just watch my YouTube videos. Yeah, I mean, I would like to start a community that is open for everyone eventually. The problem is spam. You get so full of spam so fast. Even in my current Skull community, people always try to promote their stuff in sneaky ways. I have to set my foot down and be the bad person, but I don't mind doing that to protect something worth protecting. For the brain, Notion, Obsidian, Google Sheet, or just database, honestly, GitHub. Your brain is GitHub. So in the Cloud Code routine, when you set it up, you connect it to your GitHub repo, and that can serve as your brain. It can store all of your markdown files, pictures even, et cetera. Can I hire you? I am not for hire, but I certainly know a lot of AI consultants. If you really need hands-on support, just send me an email, I guess. Honestly, reply to my newsletter with like what your project scope is and what your budget and that's like the easiest way I can refer you to somebody. I don't make money off of the referrals so I only refer somebody if like I genuinely feel they're a good fit. I could be the spam police. That's funny. Yeah, I mean we have a lot. We have like 30 admins as our spam police, you know, and they're still spam. Like people still sneak in stuff. Yeah, I was paying $400 a month for Beehive, and I wasn't using most of the features. I also generally do like that Substack has a social platform component. Like, currently, Substack's pretty fun. Like, it feels like a less toxic version of Twitter. There's a lot of really good, smart people on there, especially a lot of free AI content and education on there that's like... medium quality in my opinion, so that's good. Definitely way higher quality than like TikTok or Instagram, so I highly recommend Substack. So yeah, okay, cool. Well yeah, so my sushi arrived, so thanks everybody for joining. Check your inbox or check this link I'll post again if you haven't received the email yet. And highly recommend going through this. When you... really change your framework into like running loops, it is such a huge productivity unlock. Like you really can be doing multiple, not crazy technical things, but multiple minor things at the same time by like constructing a clear goal. giving it the way to verify its work and then letting it do its thing letting it cook as i like to say so um but yeah thanks everybody for joining highly recommend following through the newsletter and i put at the top you have full permission to repurpose all of my content this training in particular i know is worth thousands and thousands and thousands of dollars so i really hope you just try to go through it um because it's it's a it's a next level unlock for those of you who are especially if you're already familiar with skills you're already familiar with connectors and it's just it's it's about like changing your thinking of how you tie this all together um okay but thanks everybody for joining so okay i'm gonna end the stream now

Steal my best AI secret codes & automations free 👉 https://sabrina.dev
33M+ views last month with Blotato 👉 https://blotato.com

Loop Engineering, Claude Code /goal, and Agent Routines 🤯

Almost everyone uses AI by sitting in the loop. You type a prompt, you check the output, you fix it, then you do it all again. Loop engineering takes you out of that loop and hands the back-and-forth to a second AI agent that checks the work and decides the next step.

In this full live tutorial I build your first autonomous agent using Claude Code, the /goal command, and Claude Code routines. You will see the maker and checker pattern in action, where a smart model like Opus 4.8 does the work and a cheaper, faster model like Haiku independently reviews it, so your agent is never grading its own homework. Then we put it on a schedule so it runs in the cloud on its own, even when your laptop is closed.

📌 What you'll learn:
→ What loop engineering really means and why most AI users are stuck in the loop
→ The maker and checker pattern: one model does the work, a cheaper Haiku model verifies it
→ How the /goal command in Claude Code runs a task until a goal condition is met, with constraints and guardrails
→ How to build a Claude Code routine that runs on a schedule in the cloud, connected to Gmail, Slack, and Intercom
→ Real builds: clean a messy downloads folder, label every row in a spreadsheet, and auto-triage support tickets

This is for AI builders who have heard a lot about autonomous agents but have not built one hands-on yet.

⏱️Chapters:
Section 1: Loop Engineering Concepts
0:00:00 - Introduction
0:03:01 - What Loop Engineering Means
0:04:35 - Most People Are Stuck In the Loop
0:06:02 - Concept 1: Automations on a Trigger
0:07:37 - Concept 2: Each Agent Its Own Work Tree
0:09:06 - Concept 3: Connectors Let Agents Act
0:12:03 - The Maker and Checker Analogy
0:13:39 - Concept 4: Shared Memory Across Agents
0:16:32 - Use a Cheaper Model as the Checker
0:19:42 - Don't Let Agents Grade Their Own Homework
0:21:14 - The Hardest Part Is Defining the Checker
Section 2: The /goal Command in Claude Code
0:22:48 - Demo: Why Slash Commands Matter
0:24:00 - Goal Example: Sort and Triage Email
0:25:34 - How the /goal Command Works
0:28:44 - Live Demo: Clean a Messy Downloads Folder
0:31:38 - Goal Conditions, Constraints, and Guardrails
0:33:10 - Example: Label Every Row in a Spreadsheet
0:34:38 - The Empty-the-Queue Pattern
0:37:32 - Keep Your Goal Prompts Simple
Section 3: Claude Code Routines
0:39:02 - Routines Run in the Cloud
0:40:38 - Creating a New Routine
0:42:00 - Choosing Model, Trigger, and Connectors
0:43:33 - Testing a Routine With Run Now
0:45:13 - Adding Connectors in Claude.ai
0:49:31 - Real Example: Auto-Triage Support Tickets
0:51:03 - Setting Environment Variables and API Keys
0:52:38 - The Checker Agent Verifies the Work
0:55:37 - Cheat Sheet: Ship One Tiny Routine This Week
0:58:32 - Why Skill Still Matters With AI
1:00:09 - Recap of the Core Concepts
Section 4: Live Q&A
1:03:21 - Live Q&A Begins
1:04:47 - Autoposting With Claude and Canva
1:09:07 - Why /goal Beats a 'Don't Stop' Prompt
1:10:48 - Using AI for Content Generation
1:19:35 - How Sabrina Runs Agents Day to Day
1:22:34 - Editing Video With Claude and ffmpeg
1:25:34 - The Free Skool Community for Women
1:30:08 - Guardrails to Protect Your Context Window
1:31:42 - Cold Outreach: Lead With Value
1:37:30 - Claude as a Swiss Army Knife
1:40:35 - A Canva MCP Formatting Tip
1:45:06 - Closing: Loops Are a Productivity Unlock

Alternative Titles for the algo:
• Build Your First Autonomous AI Agent with Claude Code
• Loop Engineering Explained: How to Let AI Prompt Itself
• Claude Code /goal and Routines Tutorial for Beginners
• Run AI Agents on a Schedule, Even With Your Laptop Closed
• The Maker and Checker Pattern for Reliable AI Agents

#ai #claude #ClaudeCode #aiagents #automation #loopengineering

📩 I made a companion newsletter for this live with all the cheat sheets. Get it free 👉  https://www.sabrina.dev/p/loop-engineering-claude-code-goal-routines

Loop Engineering, Claude Code /goal, and Claude Code Routines