BONUS: How to Use AI Agents for Total Beginners: A Crash Course w/ Agent Builder James McAulay

32:20I wasn't in, there we go. I wasn't in the chat box. Okay, so it's going to basically go off, fire some code at Todoist and then it's going to update my task, come back and say it's done. For people who don't know, Todoist is a to-do list app. It's a to-do list app, sorry. if that wasn't. You might have said it. I was looking at the chat. So now we have, you know, a bit more detail specifically about the live stream. It's changed the priority of the task. And so theoretically now, I could sort of start my day and say to my agent, you know, tell me what is on my plate for today. Like what is due today?

32:59Help me organize my tasks. And actually, the more you start to work with AI, the less I notice myself actually opening these apps. so I'm not going to show you my actual to-do list but I use this there's a board of tasks I have here and I use it as like a shared task list for me and other AI agents so I can sort of start my day and look over in Claude I'll say you know what is due today what do we have to work on and then I'll say to Claude you know are there any tasks here that you could take on and it looks through the entire board and says well actually you know there's one here about drafting this reply or like investigating this bug um let me go and look into that while you go and do something else so you know the level i guess level one here is is claude can see what you are doing as you like as a human level two is that you can start to like delegate work to claude by asking it to take stuff off your to-do list um so so i think this is i think this is really cool I wish I could show you my to-do list, but there's just some sensitive stuff in there that I'm not going to get into right now.

34:10So yeah, so for me, I think some of the most helpful integrations that you can add are whatever you use to manage your tasks, plus whatever you use to transcribe your meetings um so if you've got something like granola or fathom running during your you know your meetings at the end of the day or you can use google meet i think zoom does this as well it's a little bit more fiddly to get like google meet transcripts into claude easily but it's definitely possible then at the end of the day you can say you know like uh go through all my meetings today and tell me, you know, what should go into my to-do list or give me a quick summary of everything I said I would do, for example.

35:02And your agent can just sort of go through, scan and then maybe drop some stuff in your to-do list. So let's put this in context for people. So this is quite literally, you have a, let's say like to start$20 subscription to this, it could be Claude or Chat2PT in this case you are using what's called Co-Work which is we don't need to get into how that works but it's just a version of Claude it's not just chat it can actually go out and take actions so that's the agent that you're working with and then once you connect that with plugins to other applications you can now talk directly to the agent in Co-Work and have it go out and take actions based on what you're asking Yeah, exactly.

35:46So like, actually, let me just kind of walk you through the connections I've got. And very briefly, I'll show you like everything that my agent is able to do. So I'm going to save this one for last agent mail with Apollo.io. This is for like enriching contacts by, you know, if I have an email address or a LinkedIn profile, I can say to Apollo, go and find out everything you know about this person. and so if someone emails me or you know joins my mailing list and I'm going to reach out to them you know to talk about a workshop for example I can ask Claude like quickly go out and you know enrich everything in my CRM about this person and and it will do that with Apollo with Figma I can ask Claude to create mock-ups for me I can ask it to create diagrams with GitHub it's connected to my website it can see my calendar it can see my drive it can see all my meetings it can see my notion I use resend for sending some emails it can see my to-do list it can see atio my crm you know this is what I was saying at the start I was saying like my whole business is plugged into Claude so that I can ask Claude anything about a customer anything about my tasks anything about a meeting that happened yesterday, and you start to come up with, you know, you start to have Claude doing pieces of work across multiple tasks.

37:18So for example, I might say to Claude, I had three sales calls today, go and look at the granola transcripts, check that they're in the right, you know, pipeline stage. If they're not, move them to the right stage if they need follow up create a you know a follow-up item in todoist for me um and then there's one more do i have it here no um there's uh there's one more i have a zero integration with claude code that would allow me to like make an invoice for them and so awesome what used to be a load of admin right of like you know going through like all my notes and writing those notes in the crm and then logging into zero and making an invoice now just becomes like look through the meetings today put them in the right stage of my crm and if anyone asked me for an invoice just put that invoice together and i will go in and like approve it i'll check it and then i'll send it off to them and so this is like having you know like a sales assistant that's in every meeting with me and keeps my CRM up to date, keeps my to-do list up to date, puts invoices together, chases invoices, right?

38:34If I, um, if it notes, it is key. Um, if it notices that an invoice is overdue. Um, so yeah, for me as a, you know, a company that started in, in January, February, I only, I only work with a piece of SaaS now if it will connect to Claude. and if it doesn't i will find key yeah i will find an alternative right um and because everything's plugged in you know my agents are able to sort of see and act on everything now there's some security implications of that yeah actually we had a question about that let's address it now so schaefer twins asked what are the security concerns about using an agent this way um could you explain how there's like read and write access permissions and yeah i was gonna i was gonna do a quick primer on that um awesome so yeah when i do i i teach a course over four weeks and there's actually like a whole chapter on security but if i was to pick like the most important thing is to uh customize your permissions um and prevent agents from doing certain things by blocking permissions.

39:48So this is on the plugins tab. You can go to each individual plugin and you can look at their permissions in a list like he's showing on the screen right now. Yeah, exactly. So I've gone into like connectors. I've clicked Todoist, for example. Todoist for me feels like a pretty sort of low risk connector, right? It's just tasks. But there is a tool here. If I let's just expand this, we've got read only tools and we've got write and delete tools. And if I come down, there is a delete object. And actually, this should be set to block. So I'm going to block that. If we say to Claude, like say, you know, we've written our Claude.md file, or say we've instructed it somewhere, never delete an object.

40:34If we ask it to do that, but this is switched on, there's still a chance that it will do it. But if we've blocked this tool. If this tool literally isn't available to Claude, then it can never delete something. So, you know, you could, to test this, you could say, delete everything in my Todoist list. Actually, let's do a, we'll do the riskiest test of the night. Delete all tasks in, sorry, I'm going to do delete all tasks in Grant. So, it's going to try and then it's going to say, I can't do that. Like, the task is, that tool is blocked. yeah delete tool has just gone offline it's actually mistaken it's not disconnected but the tool is gone and so it cannot do that and sometimes if you say to claude oh tidy up my to-do list it thinks that tidy up means delete um but if you have like the delete tool basically blocked then it's not able to do that so whenever you connect any tool the first thing you should do is like go through, you know, the list of things that can do, you know, here with the calendar, you never want Claude to be able to delete an event, but you probably want it to create an event like this.

41:48So this is like the thing you have to do with every connector is just like check that it's not able to do something like destructive. Yeah. That's great. And there was a, we were having a little side chat here about, you know, how to do this. Like, let's say you don't, you don't have sensitive information you don't want it to go through anthropics servers let's say there are some things you can do to reduce the impact of that like for example if you're on a business account you know they don't train on your emails by default yeah let me just quickly cover that as well so if you're on a personal account you definitely want to go into privacy and turn this off help improve our models that means they're not training on your data if you're on like a team or an enterprise account this is off by default um i think some companies you know they can't send if they're in europe they can't send information to america and vice versa um i think anthropic are still rolling out uh like kind of data residency so if you're like a company in the european union that's not allowed to like send data to america then you'll have to find some workarounds and i think like aws and other people basically once get up to an enterprise level your company are probably trying to figure this out or have figured this out already yeah yeah that's great and there's also we just had an interview that we just released yesterday with uh dr elena zoo from um intel and she actually created a solution to this for herself because she couldn't her company policy you know doesn't let her upload sensitive data over the cloud so she created a local email uh agent that basically runs on your computer you can learn all about that in that video.

43:29I just shared a link to it. Yeah, I think that's the future, to be honest. As these models get smaller and, you know, with some tasks like summarizing emails, you don't need Fable. You don't need like a super advanced model. You just need a cheap model that might even run on your computer. So yeah, I think that is maybe the future, sort of local AI. There was one other question I want to address before we move to the next part, which was from Evan Uho, which is if a business has 100 ,000 emails, 100 ,000 of files, etc, etc, how do you not run out of context if it needs to search through a large organization?

44:03What would be your advice on that? Oh, that's a great question. Yeah, great question. All right. Well, if you are talking about something like a knowledge base, like Notion, then there's a few things you can do. I'm trying to remember. Let me just see. I have a diagram on this, actually. Let me see if I can quickly pull it up. But you basically want like indexes explaining where all your files are. You want summaries at the top of each file. You want files linking to each other. There's a really sort of great concept called the LLM Wiki by Carpathie. and if you are like you know maintaining a knowledge base or if you're maintaining like notion you can basically ask your llm apply these principles to our knowledge and it makes it really easy for agents to sort of navigate and find what they're looking for it also makes it cheaper so that they use like fewer tokens to get what they need so that's one idea um i've seen companies get to a certain size where they start to use a search tool like glean like you know if you've got like 500 people you've got literally sort of hundreds of thousands of documents you know millions of rows of data um glean is a way of sort of speeding up search and and getting faster like cleaner data from like huge systems so yeah once you get to a certain point where you start to need like specific search tools like glean but you know in in my business probably in yours grant like my agents are able to find everything they need and I haven't had to kind of upgrade to anything like this yet.

45:45Yeah, the way that I do it is just basically, you know, whether I'm using, I've been using ChatGPT for work tasks lately because that's what our company's on. And so they have a tool there called workspace agents, which actually takes a lot of what you're explaining to us and puts it together into a single kind of like, like user interface, I guess. but absent of that you can use the project instructions which is a version of the system instruction which James explained at the beginning which is just like a bunch of information that tells you where all of the files are and what they're for and when you might need to use them and then we're going to get into skills but you can also put that information in skills so basically skills will know what to reference where sorry go ahead yeah it's the perfect segue actually because I'm literally about to move us on to skills.

46:34So we've covered, you know, what's an agent and the importance of context, starting to set up like a second brain so that your agent sort of knows everything about you. We've talked about MCP, which is basically just a connection between Claude in our case, but basically any agent and any piece of software. So let's talk about skills. I have taught 400 people. And whenever I do this course, I ask people to kind of tick the concepts they've come across before. And I'm always surprised how few people have actually discovered skills before. I think it's usually about one in four people that I survey have actually heard of agent skills or even used agent skills.

47:19I usually have a clip from, let's see if we can find it uh you know in the matrix when neil basically learns kung fu and he downloads the abilities to do kung right i think this is the perfect analogy in this you know neil is the agent he's downloading all these abilities and skills and then he gets dropped into this sort of room with morpheus to practice his kung fu um there is actually a feature in claude that helps you create new skills and then test them out in this kind of dojo with morphine um so let's let's get into skills i think skills are amazing um basically a skill is down sorry oh i did not know that how long sorry how long has it been down just the minute while you were explaining okay damn did you see the matrix i was showing the matrix no no we missed it oh no let's run it back let's run it back let's run it back can you see this yeah yes yeah so so he's downloading all these different types of martial art right and it's happening nearly instantly and then he gets into the dojo with morpheus and this is where he kind of like refines the ability and like tests his new skills against morpheus um so i think a skill is really straightforward It's just a saved instruction.

48:44It looks like a huge prompt that tells Claude how to do something. And instead of you keeping all these prompts saved in a document somewhere or keeping a folder of prompts somewhere, you just install these as skills. And then you can ask Claude to do this anytime you want. So skills have a few cool features. they they can trigger themselves like if I have a skill called I don't know if I had like a summarize this skill and I said to Claude summarize my email it would say oh you have a skill for that I'm going to use the summarize this skill I'll show you in a second but there's a way to basically refine skills so that they always produce good results and so instead of you kind of freestyle prompting, asking Claude to do the same piece of work every day.

49:41You can like package it up as a skill and you know that you're going to get like a pretty similar result every single time. Once you package it up as a skill, you can sort of optimize this skill. You can refine it. It kind of keeps getting better. And then crucially, you can share skills with teammates. You can share them with customers. I know some companies that like, you know, give their customers an agent skill that helps to install the software for example um wow that's smart i like that it's really smart yeah so basically like you package these up and they're they're incredibly valuable um so the way you use them is you know you go into the plus you go into skills and i've actually got like a bunch of got a bunch of skills uh i've got one you know that looks for tasks in a gorola transcript and then drops them into Todoist.

50:34I found one on the internet and I'll show you where you can find skills in a second. But there's one here called Humanizer that sort of tries to strip AI signs out of your writing, like the M dashes and the, this isn't just this, it's that. It's key. If you're posting on LinkedIn, please use this. Yeah, yeah, yeah. Yeah. The AI swap button on LinkedIn, I think it might be working. I swear I'm seeing less slop now that they've launched that button. That's amazing. Yeah. I was building an agent called Jamie. And so I have like a tone of voice skill for Jamie that like makes sure that whatever I'm writing is sort of like on brand.

51:11But you could also have a tone of voice skill for yourself. I've got some security skills. I've got a skill for product requirements. You know, these are all like different tasks that I do, you know, weekly or every morning or every now and then. and you can either create them yourself or you can find skills online so i'll show you quickly i think the best place to find skills is this website called skills.sh and this is slightly more aimed at developers uh but uh there's a load of good stuff in here they've got one million skills here. And basically I can search, you know, marketing and we've got a whole bunch of like popular skills, you know, with a hundred thousand installs.

52:01This one's called marketing psychology. And this guy, Corey, Corey has like packaged up a bunch of marketing skills. There's a summary here, but basically this says to Claude, you're an expert in applying psychological principles and mental models to marketing. So your goal is to help understand why people might buy something, how to influence their behavior ethically, and how to make better marketing decisions. And then the skill just goes through, you know, a whole bunch of like ways of thinking first principles, jobs to be done, the circle of competence, Occam's razor, it kind of loads all this into Claude and says, Okay, we're going to think about why someone might buy something like use these principles to help us like improve our thinking so this is a bit of a conceptual one but then you also get quite tactical ones like there's an seo audit skill and this skill will go through your whole website and it will look for you know any context that you might have and then it will sort of uh basically check whether you're getting good like rich results in Google.

53:04It will see if your site is like easy to crawl by Google, easy to index. Like, is it fast? Do you have good links? And it runs through this whole process, right? This is telling any agent to follow this process and like audit your website from an SEO point of view. So skills are incredible. And once you discover them, you can just like drop them into Claude immediately. And you give Claude these abilities that have been refined by someone else but I think they're most helpful when we create our own skills and the way we do that is we go into the plus we go into skills and then we say uh oh interesting okay well the way we do that is skill creator which is kind of hidden.

53:51And so I could say I want to create a skill that looks through my Gorilla transcripts at the end of the day, pulls out commitments I made, and summarizes the key topics before putting together a report. okay so what this is going to do is walk me through a process uh for creating the skill it's going to interview me it's going to ask me you know about what what type of summary i want do i want bullets do i want paragraphs like it's going to work with me to create um a skill it's going to show me a few sample you know summaries and then it's going to ask me for feedback and I'll say, oh, this report is way too short.

54:50This one's way too long. Looks like your screen dropped again. If you could pull it back up. Sorry about that. Let me... I think it's just sometimes it gets a lag spike or something. Yeah. Can you see this now? It's loaded in. Yes, we're good. Yeah. Okay. So I have actually created some fake calls. uh you know an example is like a fake transcript where you and you and cory were just talking about like your goals or like planning out the academy you know this is all just like fabricated but you can imagine oh no it dropped again it dropped again no way no way sorry about this um let me just close some stuff down and see if i can uh let me see let me try that's hilarious i wonder how close it is to the actual transcripts we would have um so i just created these so that i didn't use my own granola calls uh but it's like you know planning the academy planning your goals so it's going to interview me i'm going to say like let's make a new skill i want you to yeah let's save a report file uh should you write anything into todoist uh actually yeah that sounds good let's do that and so gradually it's refining this like piece of work with me we may not have time to go through the whole process but it will then show me like three examples and ask me what I think I'll give it feedback it will give me more examples I'll give it feedback and then eventually it will say here's our skill you can run this anytime you want and I'm pretty sure you're going to get good results um if you don't use skills then you know what you're going to be doing is effectively going to Claude every day and saying like look through my granola and find my commitments and you're going to get different results every time a different format so having the skill just makes this you know like really uh predictable and and really um sort of like uh deterministic like yeah you you get less of a sort of freestyle with claude and you get a more reliable um output how do you ensure the quality floor like how how do you make sure that it meets your standards every time what's your advice there um let me just see if i can get to the end of this and then i'll i'll explain exactly how i do it uh

57:09cool so once claude has had a goal at creating the skill it's going to ask me for feedback and my advice is give claude as much feedback during this step as possible like when i'm creating a skill i i'm not joking i might spend like an hour to an hour and a half making a skill if i think that it's something i'll use every day or every every week um and the more feedback you give claude you know this part of the summary is too verbose uh you picked out something that was actually you know quite trivial only focus on like bigger you know bigger strategic items if you give it all that feedback eventually it starts to give you a report that does match what you want um so yeah my advice is to um is definitely just to like take this process like the skill creator process quite seriously and like spend some time on it um i also sometimes if i'm creating a skill for example i've actually created a really helpful like morning stand-up skill that plugs into a weekly retro skill and so each of these you know i run it in the morning or i run it on a Friday afternoon and it basically puts like a report into my second brain about what I got done that week and sort of asks me a few questions and so I went through this process of skill creator and I got to a point where I thought like okay the examples I'm seeing are quite good but why don't I simulate these skills on like some previous days right go and look at you know the metrics for the previous days go and look at my calendar from previous weeks and so then I'll say to Claude, spin up 10 sub agents and simulate this skill across the last 10 weeks.

58:51And then look, look through all of the outputs and tell me whether you think this would be genuinely helpful. So that is so smart. So yeah, basic or if I'm, you know, this, the skill that I shared, if I go to the skills page that I mentioned, right, agent accelerator slash skills, one of them is called like set up my second brain I've got one called like uh multi-agent etc when I'm giving away skills to other people I'll usually say to Claude like simulate 10 different people going through this and tell me whether you think they would get stuck anywhere or whether the interview might fall down so yeah it's a blend of like giving it a lot of feedback yourself and then asking it to like simulate the skill being run multiple times um so that it can sort of reflect on whether it's it's going to work yeah awesome that is that is so helpful so being able to tell it to go and simulate it and test it up front because the way that i do it is i will just then start using it in practice but i use it pretty manually i would say my process is still pretty manual these days um in terms of me chatting with the agents and and you know i'll set up scheduled tasks for it but then i'll go in and i'll you know continue continue the work to finish it myself so yeah i will update the skills as they go i will say like oh hey this is wrong let's make sure we never make this mistake again and i'll update it in the chat but i'm not doing this level of um detailed uh create detail up front in the creation process and i think that's brilliant i think this is yeah this is also like a feature that people aren't aware of in co-work is that you can ask Claude to spin up sub-agents.

1:00:36So you have to ask for it explicitly. But let's say simulate 10 fictitious people. So fictitious people going through the morning stand up. Oh, my goodness. Going too fast. Actually, no, the weekly retro. There we go. Weekly retro skill using sub agents and reflect on usefulness of the skill. that's that's a bit of a broad example but i just want to show you that like basically claude is gonna spin up sub-agents that each take on like the persona of someone completely new and go through this skill for you and go through the whole process and then claude the kind of master agent will reflect on the 10 responses and we'll say oh okay eight people had a good retro two of them like this didn't really help them and that's because we asked a bit of a vague question for example um so cool so that's that's skills um and now we're on to the most exciting part which is setting this up to run like in the background uh while your computer is off so let's get into this because this until like a couple weeks ago was quite fiddly um in cloud there's like four ways to run tasks on a schedule.

1:02:00So let's focus on co-work because I've spent most of today talking about co-work. We have scheduled tasks in the cloud and we have scheduled tasks that run on your computer. Now, this is the feature that's been around for like a few months. And if you've used this in cloud co-work, you've probably been using local tasks. This has access to all the files on your computer. So if you've got, you know, your second brain or your files sitting on a folder on your computer, then you need to leave your computer on. But you can say to Claude, every Friday at 9am, I want you to look through, you know, all my files, all my MCPs, you know, granola, whatever.

1:02:42And I want you to send me an email that summarizes what's coming up today. and this is just a bit fiddly like i don't think many people want to leave their computer running 24 7 um so when i say to people you know if you're going to set up a scheduled task do it at a time when you know you'll be working like if you typically work like 9 30 till 5 30 set it up for like 9 45 like your computer is probably going to be on just make sure that cloud is open and it will run but they recently launched cloud scheduled tasks and these run like on anthropic servers your computer can be off but they cannot touch your local files because your computer is off so it can access you know like notion or granola it can see your skills but it cannot access these files on your computer which leads me on to actually notion and why i think keeping a second brain in notion might actually be more helpful than keeping it on your computer because if claude has all the context it needs in notion then you can actually set up a cloud task and you can say you know look at my goals look at my background and i've duplicated everything i showed you right Like the goals for 2026, like a million readers by end of 2026, getting 5 ,000 views per live stream.

1:04:11I don't know how many we've got now, but I'm guessing it's, you know, 100 ,000 easily. You know, so like everything that I showed you as markdown files on the computer could also be a folder of Notion files. And then because it's in the cloud, we can go into Cowork and we can go to Scheduled Tasks here. And there's this new feature, you know, run a task in the cloud. And we can set this up manually. And we could say something like, look at my, I think I have a prompt. Did I have a prompt? We could say, yeah, look at my meeting transcripts for the current day and then assume we have a skill called, I'm going to just use this one, granulated todoist skill to pull out commitments and save them to todoist.

1:05:10Email me using resend, and I want this email to come to agentaccelerator.ai using resend with a summary. look at my goals in notion to calibrate the uh the task again i would spend more time writing this prompt but this is actually able to go through all my meetings drop stuff into todoist send me an email understand everything about me from notion without touching a file on my computer and so i could set this up as a cloud task i could call this like end of day uh you know like chief of staff report. I can say, you know, you're actually, to the purposes of the demo, I'm going to say like, you know, you're allowed to do all of this.

1:06:02You don't need to ask me for permission. I'm going to say this is a Sonnet task. I'm going to say let's do this every day at 5 p.m. And then I'm going to leave this off. So this is like, you know, run on your computer is off and actually it will run in the cloud. And so if I then, let me just, I want to try and show you this working because this is like the kind of final demo, right? If I say to Claude, meeting transcripts, this is a demo. Meeting transcripts are in Notion here.

1:06:52And I also want to make sure it saves to grant project in Todoist. Okay, so I'm going to save this and see if this works. So we've got our end of day chief of staff report, and I can test it out now by hitting run.

1:07:13Okay, it says it started.

1:07:18so if I come in here does it show me where it has the task that's our simulation yeah that was it this is our simulation by the way scheduled right at the top yeah right up there so if you go on it's scheduled yeah oh here it is it's running here at 708 so it's looking through all the tools it needs and it's basically gonna it's found the thing in notion it's found the grant project and todoist well have i got all the connectors switched on i'm going to need to switch on resend if it's not on there we go um connected turned on and so it's basically gonna like run through and then hopefully spit out a report for us and a bunch of tasks in todoist based on these transcripts of meetings that I have, you know, fabricated.

1:08:11It says it couldn't run because Notion and Resend weren't authenticated. Okay, that's annoying. What happens if, what do you do if that's the situation? What do you do? Yeah, I think the problem is that it's, yeah, it's trying to use a granola, it's trying to use a granola skill instead of, so it's trying to like hit granola when actually I'm trying to do a demo with some Notion transcripts. Right. But I basically, I actually, I set this up and I actually got it working and I was really impressed at like how easy it was. And I now actually have an agent. I can't show you again because it's summarizing all my real stuff.

1:08:55But every day at 6 p.m. it sends me a summary of the meetings and it says, you know, here's five things that you said you were going to do and I've dropped them into Todoist for you. So that's like one example. I have an agent every Sunday that does an SEO scan of my website and, uh, tries to basically, uh, write code and submit pull requests to get hub to fix any issues or like speed up parts of the site that have maybe slowed down. Um, I have a sales agent that wakes up every morning at 8am, sends me an update on the pipeline, tells me about the calls I have coming up, uh, gives me like a sort of report on anyone.

1:09:33That's like a first conversation, gives me some background on them. So basically... Oh, I think you unplugged your microphone there.

1:09:46I can't hear you.

1:09:51Maybe if you click the audio button, unmute, unmute. I think it's a very sensitive microphone. Sorry about that. So yeah, what I want to end on then, and I haven't been able to show you the live demo, but I promise it works. We've actually done a version of this before. And so people can also watch that video and see basically how the scheduled task works. We have not done the cloud one. So that's pretty cool. The cloud one is cool. And again, you just need to make sure that everything it needs is in an app that has an MCP, like Notion, not on your computer. Uh, but you know, a year ago, if someone had said to me, I want you to like create an agent that does all these things, has all this context, like wakes up and does stuff, you know, at 7 PM every day that I would have had no idea where to start, but hopefully I've like kind of demystified that today.

1:10:49And I'm sort of showing you that really, if you have all the right context and all the right MCP connections with some skills, then it's actually quite easy to start setting up these agents that kind of proactively try to get work done for you yeah and it's doable on the max plan which is 200 it's kind of doable on the 20 level but you have to keep it pretty simple scheduled tasks are what will will run up the the usage limits because for people who don't know anthropic has essentially rate limits where you can only run so many tokens per hour and per week so you got to keep that in mind um yeah james we still have a couple hundred people in the chat are you down to hang out for a couple minutes oh yeah i'd love to i'd love to answer as many questions as we can yeah absolutely awesome awesome uh but did you have any any final thoughts to wrap us up here uh i actually had one more thing i wanted to show um let's do it yeah let's do it uh let me see and then we'll do uh we'll wrap up the rest of the questions and answer everybody's let's do that yeah let's do that okay so I wanted to today I wanted to like show as much sort of hands-on like inside Claude stuff as possible but there are a few concepts that I think are quite important to teach as well um when we move from these like reactive to proactive agents um and you know I've got some examples here of like the sales agent for me that's sending me the briefing and updating the CRM for me the SEO agent that's like scoring my website and actually like writing code uh i've got a cfo agent every month that like looks across my business and sort of tells me you should cancel this software or like oh hey cloudflare was like a hundred dollars why is that um i have this like retro agent you know all these agents um i think of these like four sort of levels of proactive agents um so level one is an agent that gives you information right and just tells you like here's a summary of everything that happened level two is an agent that gives you the summary but also makes some suggestions so if the agent knows your goals and it has all the context on your business then it can say you know i've looked through your pipeline and i actually suggest you focus on these three clients today because they have like the highest potential revenue they've got a really good fit with your icp and so now we've got an agent that's like giving us helpful suggestions instead of just summarizing information the third level is where your agent starts to actually draft work so when i for example have an agent saying you know you had five calls yesterday here's the summary i suggest you follow up with these two people and by the way i've actually drafted an email for you to send to those two people because i've got the whole transcript i've got everything i need in the crm um here's the draft here's the drafted invoice all you need to do is basically approve this like that is way more helpful than just the agent telling me what happened right um but it's surprisingly difficult to get to this this stage of like drafting the work and in the final level is when the agent can sort of reflect on its own performance and so if it can look at you know if it's a daily agent if it can see every briefing and it can look at them and say actually i've noticed that i'm mentioning the same thing every day and james never replies to it then it can like tweak its own instructions it can like edit its own clod.md file and get better and sort of compound every single day so wow the trap that people fall into the trap people fall into is just building a briefing and then ignoring it after like a week but if you actually get an agent to like do work for you and like improve every day this is like a very like sort of exciting and thrilling place to be oh totally yeah i think this is this is so key i feel like that's a really good way of framing it out because a lot of times when we think of a scheduled task it does feel like a one-off thing but if you can take it from you know one-off to know it's actually you know not recursive but it's to the point where it can look at how its performance and change and improve it I mean you're off you're off to the races there yeah yeah and like agents improving themselves is, you know, like I just showed earlier the agents spinning up 10 sub-agents to kind of simulate a skill and then improve the skill.

1:15:15Like if you really wanted to burn some credits, you could have it do that on a schedule. You know what I mean? Like you can have it do that quite frequently. I don't advise that, but you could. No, me neither. Shall we answer some questions? Let's do it. So there was a couple that we put off, but we got one from Learn Promptly AI. I wonder if you customize the cloud instructions for your profile slash account to create skills, if that'll work. Do you have any insights into that?

1:15:45James McAulay:Or do we need to reframe that question?

1:15:57James McAulay:Oh, oh, oh, your mic. no way not again not again um sorry uh if your claude md file you know tells claude to always be quite uh brief um then that might influence the skills unless the skill itself says this should be like 2 000 words long so yeah i guess those two files do sort of interplay with each other yeah i i interpreted that question slightly differently like could you edit your instructions to have it create and use skills that's kind of what i thought with i thought that was good oh i see um yes you could add a line to claude.md that says whenever you see me doing repetitive work uh um suggest a skill that we create um that actually that reminds me of something yes that I shared last week.

1:16:56I'm going to share my screen again. Where is Riverside? Here. So I have a skill called Workspace Review. And every Friday, I shared this on LinkedIn. And basically, every Friday, it looks for ways to improve your context. It looks for skills that you could create. So if you're doing the same piece of manual work, very repeatedly it sort of like says why don't we make a board update skill uh if you've been giving it important information in the chat that's not saved as a file it will like basically solidify that as a as a markdown file sometimes it will suggest like connectors if it sees that you're like pasting in images from a platform it will say hey you keep pasting screenshots of your crm did you know you can connect your crm and then it will actually look at cloud md every friday and sort of say uh you know you keep asking me to be more brief i'm gonna update claude.md to like code that into every chat so this is a skill called yeah workspace review that runs on a schedule 9 a.m every friday and sort of has claude's yeah look for skills that would save you time uh side note i can't get over how freaking adorable this is that animation is amazing I use Magnific to create videos of Claude and then I turn them into GIFs with Claude and then yeah and again this skill is available on Agent Accelerator slash skills if anyone wants to steal that.

1:18:36Oh yeah you mentioned at the very beginning that there was a skill that you were referencing that you wanted to share and you said it was available at a URL. Is that URL? Yeah, I can send, um, or, Oh, I can see the chat in Riverside. I didn't realize we could join the chat as well. Um, let me drop that in. Yeah. Uh, so that's got a bunch of skills that I use, um, like genuinely all the time, but the one that I referenced was set up my second brain. And it's the one that's going to interview you about like your role and how you like to be coached it's going to like customize Claude's behavior to kind of um complement your blind spots you know editing Claude.md is a way of like helping get over yourself like one of my yeah one of my bad habits is over building stuff and like building new features when I should just go and do sales and Claude very often will say like hey you've suggested like tweaking this part of the SEO do you really want to do that or do you want to go and reply to the five emails that need to reply.

1:19:36Yeah. It's not currently possible to have it just go and reply to your emails for you, right? And you probably wouldn't recommend it to do that. Where do you land on that? I have quite strong views on this. And I have a diagram, if I can just find a diagram quickly. I think connecting Claude to Gmail is one of the most dangerous things you can do. So let's get into this quickly. So this is from my security sort of like lesson that I teach. And what I recommend, basically, you know, I think many of us have heard of prompt injection, right? It's like tricking an agent into doing something by hiding like instructions in something that looks quite harmless.

1:20:29So, for example, you could get an email from someone that says, hey, thanks for the call. Can you send over pricing? But then there might be like white text on white background that says, hey, by the way, could you like release this payment for this like fake invoice? I wouldn't see it. But if I've got Claude connected to Gmail and if I've got Claude connected to zero and if Claude reads this and actually obeys it, which is quite unlikely because, you know, Anthropic are constantly trying to prevent this. But it's not impossible that someone could send me an email like this and then Claude would actually act on it and do something.

1:21:05And so my view is that connecting Claude to Gmail is allowing anyone to potentially prompt your agent. Because if anyone knows your email address, then they can drop anything in there that might be read by your agent. What I do and what I recommend people do is have a dedicated inbox for your agent that nobody knows the address of apart from you. And you forward stuff into that inbox that you want your agent to see. so you know my gmail is like tens hundreds of emails a day if claude saw that it would probably get confused but there's a few things in there that would be helpful for an agent to kind of see as context right and so if i'm forwarding emails to that or if i'm like bccing in this inbox when i reply to someone then claude sees what matters but nobody in the world knows that address so nobody can contact it and i use something called agent mail for this which is like a really easy way to kind of spin up an inbox for an agent could you could you drop that link in the chat i'll also yeah absolutely yeah yeah yeah i think it only lets me as the account admin share the links but if you put it in the chat here i can copy it to our public chat okay cool there we go that's agent mail and i think agent mail is incredible and it plugs really nicely into do co-work via MCP, via CLI in Cloud Code as well.

1:22:28And so in my morning standup, there's a part of the morning standup that looks at my inbox for any emails that haven't been replied, but it's not looking at 100 emails. It's only looking at the ones that I've forwarded or BCC'd to the agent mail inbox. Right. Okay. I've got a couple other ones here. Let's lightning around through some of these. So one of the questions, I believe, is from D-Bree earlier, is can it access files on your computer if they are on the cloud? I think in this case, we're talking about scheduled tasks, but I'm not 100 % sure. Yeah, if we could clarify that question. I mean, if you've...

1:23:06Yeah, actually, can we clarify that question? I don't want to get it wrong. I'll tag them if they're still here. Cool. Okay. Okay, so then the next one, Evan, uh-oh, had two more. Number one, what's the best way to run Claude code remotely when on the go? I can talk a little bit about this unless you have an option you like, James. If you have a powerful desktop at home but travel. Okay, a slightly different answer for that. And then number two, you said to use Fable. How do you pick Fable low effort versus Fable high effort or even Opus 5 max versus Fable low? Basically, what's your advice on effort and model use?

1:23:48And I've answered for both of those as well, if you don't have one. Let's do the first one together. So what were you going to say about running Claude from home? So the way that I use it, and this is based off of Boris' journey, comments he made a couple months ago. So who knows? He might do something different now. Boris is the creator of Cloud Code or one of the creators. And so when he says this is how he uses it, I say like, okay, that's all I'm going to do. Same. I think they might now do most coding in the cloud. But at the time, he said that he starts the project locally on his computer in the cloud desktop, uses auto mode, and then remotes into it.

1:24:34So he's actually coding on his phone with an instance that's local on his computer. So it's a little bit confusing to visualize, but basically there is, you start at local and then you use what's called remote control, which is a toggle that you toggle on. What were you going to say, James? Yeah, I have a few answers to this. I did a lot of research to this a few months ago because I was going on holiday. I didn't have my laptop, but I did want to be able to like access this agent that, you know, knows a lot about me. And so at the time, um, I used this GitHub repo, where is it called Telegram bridge, uh, for code.

1:25:20And this is like a really secure way of, um, basically like messaging called code, you know, while it's running on a computer through Telegram. And you're basically piping all your messages from telegram into the terminal which is running the code and then it's coming back and telling you what happened and i made this diagram that shows how it works right i'm on my phone i send a text or a voice note it goes through the telegram bot api through this bridge this guy made and then into my session claude does a bunch of stuff comes back and i did a load of like security work on this to make sure that only i could access it it just felt it felt very fiddly though and like it took a lot of like hacking together right um i agree that i think the the easiest way now to do this and maybe i can show you this is like how i use cloud code so i just use it in the terminal and you just want to type remote control and i'm going to enable remote control and then if i go to the cloud code app and go into code mode here we go my session is open and yeah wow it's worked flawlessly actually um if i stop sharing my screen um and i hold this up to the camera so that is actually the session that i was just having with claude um and i could just keep going as long as this computer is still running so yeah that's i think if that's how boris does it that's the simplest way to do it yeah yeah the the and there's a ui to do that in the app as well.

1:26:55So like literally you just start the chat on local versus cloud and then you toggle it on. So whether you're using the app or using the terminal. As far as the effort versus model, how do you handle that? Because you're doing a lot of agents, a lot of scheduled tasks. So I assume you're very rate limit and token conscious. I'm a little more manual, so I'm a little more flexible with that. But how do you approach it? Okay, a few things on this. when you use skill creator as well as trying to give you the best possible skill it tries to make you the most token efficient skill so it does actually measure as it's like going round and round it measures um like how many tokens you're using and tries to optimize that i use fable or opus to create the skills and then i use sonnet or maybe even haiku but usually sonnet sort of medium effort to run the skill because if a skill is quite simple like you know go through granola pick out summaries and drop them in to do this that doesn't take like a rocket scientist's brain that just takes sort of relatively basic llm so where i can i will like specify a sort of sonnet low to medium effort um and then day to day i'm using opus or fable on sort of medium to high unless i need something done very quickly and then i might drop to sonnet or drop the effort down um and i have been playing with like the chinese open weight models like kim ek3 and glm 5.2 and actually i was inside your system yeah that's fascinating it's uh it's funny you ask i i've got a video coming out on this uh next week on youtube so i've started putting more effort into my youtube videos and i've got an editor and this is my first proper video coming out you guys basically i plug it in the neuron i would i would love to thank you let me show you what i've done so i use open router and for anyone that's not aware open router is this like incredible platform for accessing like all models and you can access you know like anthropic models open ai models meta models and all of the like chinese models from like deep seek quen kimmy glm and there's a way of um If you look at the OpenRouter Cloud Code docs, there's a way of switching out, inside of Cloud Code, there's a way of switching out your, like, Anthropic URL, right?

1:29:24So it asks you, like, your base URL. They suggest that you switch out for OpenRouter and then everything goes through OpenRouter. But the problem with this is that that switches your Anthropic, like, your Cloud usage to API billing and immediately it's really expensive. So you don't want to do that. I've come up with a workaround where Claude stays on the subscription, but Kimmy and GLM actually go through the API. So I've created some custom functions. And now if I type Kimmy, we should see Moonshot Kimmy K3 API usage billing. But if I type Claude, then it shows my Claude Max subscription. so this took some hacking and like when i release the video i'm just going to release a prompt that you can sort of copy and paste and then it'll do it for you but what i've been really impressed by is that like um glm is is rapid like glm is really fast and it found stuff in a piece of work yesterday that fable missed um so yeah i'm like you know i'm i'm really impressed by this i'm really impressed and i think this is maybe like what i'm going to spend some more time exploring is you know like how far can you push these cheaper models yeah no that's awesome and that's so i can't wait to watch this video i'm definitely going to apply that um i would say so a good framework for people and anthropic released something not that long ago maybe a couple weeks ago where they tried to explain the difference between model and effort and what when you would want to use one you should think of effort as like the time amount of time that it's going to spend on the task.

1:31:01So if you're asking it to do something that really would take you or anyone else a lot of time and also a lot of context, you probably want to increase the effort and increase the model. So that's kind of the framework that I use, where if I'm like, hey, I'm going to have you look at my entire code base and, you know, draw some conclusions. Fable, you know, as high is as high as I go on Fable. Opus is five, I crank it to max. And that's because I heard some advice that X-High and Max make Fable kind of lose its mind a little bit. I've seen that. Yeah. And I've also, there's a feature they have called like dynamic workflows where suddenly Fable wants to spin up like 50 sub agents and it just burns your tokens.

1:31:46Don't do that. So yeah. Yeah. I think the one thing, the one thing that I like about it, I think when Fable came out, the one thing I thought was quite cute is if you type effort in the terminal, like their x high mode sort of like makes the whole thing if i do effort here um ultra code yeah i think this is so cool right yeah it's not even fable that's sonnet as well and that's going to put it to x high with workflows and when i did this it literally spawned like 60 sub agents and i was kind of like stop stop stop please that's way too much um yeah i agree i definitely agree with that yeah the the The caveat I'll give to that is I do think if you have, so the way that I've been using it is Fable is like my planner.

1:32:31So whenever I'm talking about anything, hey, we're going to do this new initiative on the project we're working on. I'm talking to Fable. I will have it write a detailed technical specification. And I'm like, so you're not just coming up with what we should do. You're coming up with how we should do it. And then I have it farm out subagents to do that. So my main chat is still Fable high, but all of the work is being done by Opus Max or Opus or Sonnet Max. Yeah, I do something very similar. I have a skill for product requirement documents, and that can, again, take half an hour to an hour. Then I will run a skill called multi-agent refine, where it spins up four or five sub-agents, It's usually like a model down, sometimes the same model that will like critique the document, the plan from different points of view.

1:33:23And then sometimes if I'm really paranoid, I'll get like codecs to take a look at the plan. And then when I'm sure the plan is rock solid and I've thought of everything. Yeah. Spin it, spin it up with like sonnet sub agents to kind of crank through it. Yeah. I love that. I need to apply that because I'm very much a let me get a version of it and then see all the problems with it and then go back and fix it type of person. but I think to level up my game, I got to do some of that. Okay. Real quick. D-Day Bree responded and they said, yes, scheduled tasks. You mentioned keeping files on Notion, but if you have files in the Google cloud, are they accessible in the context when your computer is off?

1:34:03Good question. With Claude, and someone correct me if I'm wrong on this, I still think that you can only read Google Docs with their integration. I don't think you can edit Google Docs. So for example, if you've got like, you know, sort of second brain set of files in Google Docs, right, about who you are and you want Claude to read it before doing a scheduled task, I think it will be fine. But if you wanted it to like create a new document every day or summarize, you know, every Friday, create a retro document, I don't think it can do that with Drive. but it can definitely do that with Notion.

1:34:40So the Notion MCP is like very, it used to be kind of bad and now it's all right. But the Google Docs, they have, yeah. I think Notion wants to become everyone's second brain. But I think Google Docs and Claude, for some reason, just aren't friends. And maybe that's Google's like one way of clinging on to Gemini. They have two ways of clinging on to Gemini. Number one is that, what you just described. And the second one is you can't use anything besides Google Apps in Gemini. it's crazy i have people come to me you know to do this four-week program and like they usually abandon gemini in the first week if they're using it because you just can't connect anything i don't know what they're doing over there um perhaps they have a new launch coming out today i heard rumors of that i'll have to check after this stream but yeah i think they're confused i think they are and i think like yeah whenever they launch something they're like oh so what you want to do is you want to go to our ai studio and then you want to open anti-gravity vertex whatever and then find our flash omni pro mod like just simplify it just yeah yeah yeah totally no i think they should just make ai studio the app and get rid of all the other ones or make them plugins in ai studio it's it's that easy and they'll win or maybe not win but they'll be competitive okay i have a couple more um real quick there's a good correction from g hogarth by the way Apparently, Claude can write MD files to Google Drive.

1:36:05So I didn't know that. That's great. Yeah, I was going to say, actually, on that point, that our company created a custom connector, essentially, so that we can control who can access what. And that's a good recommendation if you work at a larger company. You don't have to go through the traditional plugins that are in there. You can actually create your own, and you have a little more control over it. And we can read and write to Google Docs. But this is on ChatGPT is where I use it. little caveat there yeah good good note there so this was from mobility fl who if they're still here they're a real one they came in at 9 30 and asked this question and they said number one how to determine which agent is good and for what task i think we kind of covered that um how often do you reevaluate i don't know if you want to address that um they have two more questions you want me read them how often do i reevaluate like different agents or different skills i think it's like what agent you're using for what task how often are you reevaluating that so some people think of an agent as like some people have like 10 different skills right and they call that 10 different agents like some of them are like i've got my chief of staff agent and i've got my product agent and like the way I think about it is Claude Code is my agent and it has a bunch of skills and so I am very happy with like my Claude Code setup whenever a new model comes out I will sort of like take it for a spin sometimes you have to like change your skills sometimes you have to change your Claude.md like Opus 5 for example doesn't need as much hand-holding and like strict instruction as 4.8 um so yeah i feel like i'm constantly sort of trying to get the best out of the newest model but in terms of like the agent that i use i like the clod code as a harness is great and then like i mentioned i'm testing out glm and kimmy but you know the clod max subscription is so cost effective that i'm just going to keep hammering that until they up the price it really is i was pretty impressed with myself last week i was able to get 100 model usage and 100 % Fable usage.

1:38:16Well done. Nice. Thank you. Thank you. Usually it's one or the other. Yeah, yeah. The other question from Mobility was how to manage and see the tokens you use so you can optimize. There is a feature called usage inside of the app, but I don't know if you have a better solution for tracking. There is, yeah. So in the app, you can use usage and co-work and cloud code in the desktop app. they both will sort of like show you a little dial where they'll sort of warn you uh i've installed a really nice plugin in the terminal called code code hud and so this shows me the model i'm using the context of this conversation and my usage that resets in one hour um and so if i if i like sort of switch through these different you know these different chats like i can see the context of each one so i find that quite helpful but um honestly like i've been using cloud max 5x for like nine months and i very rarely hit the limit um so yeah i i that i'll drop the link to that hgd plugin if anyone wants that but um oh yeah drop it in the chat and i can share it yeah yeah um and then they had a question about edge agents edge ai models i think that's outside the scope of what we're talking about today.

1:39:34But do you have any? What do they mean by Edge? Edge would be like, you're running it locally on your computer or your phone. Oh, yeah. Right, right, right. What the best ones are for that. Yeah. Yeah. I don't know if there's an UX there. No, not really. That's not really something I've had to think about. Like, if you're building a hardware startup, right, then go and find a really small weight model. but like i i don't have much to offer on that question uh the thing is like this changes all the time but the ones that people are really happy with right now is gemma um i think it was gemma 12b was was like pretty good for the size um there's like some really cool small ones that just came out you i advise you read the neurons around the horn digest to keep track of those that's on our website um and uh and then allegedly you can use things like deep seek and kimmy k3 on your computer at like two tokens per second i don't know who is running this or how is it even worth it like two tokens a second i mean i went and explored yesterday i thought like a mac mini 64 gig ram and i was like okay it's time to run kimmy and then i looked at the size of it and i was like no way it's just not gonna run but gemma gemma would run on that machine but like Kimi K3, no way.

1:40:54Yeah. Yeah. And then there was a new Quen that just came out. I think it was Quen 3.8. And I think that one might be, people really like the Quen models, especially for local coding, but they're still not at the level of like Fable or Opus yet. Yeah. I honestly, like three months ago, I thought I was going to go down this like local model route and I bought the Mac Mini and I thought, okay, it's time to kind of like get some free, you know, token consumption. I think that just pushing stuff through OpenRouter and spending pennies and testing out the models in the cloud is way easier than installing all Lama and waiting three hours to download the file and set it all up.

1:41:34I think OpenRouter is amazing and a really easy way to just test new models as they come out. I agree with that 100%. Let me share that GitHub in here. I will say with the caveat that I am very bullish on local. I think we should have Opus or maybe even Fable level of AI locally on our computer. And that's the future. And we'll eventually get there. We just talked to Intel's Dr. Olena Zhu. And she basically said, you know, if you look at the chart, if the trends hold, we could have Fable level AI on our laptops within two months or two years. Sorry, not two months, two years. Yeah, I don't doubt that.

1:42:10And I also agree. And like, you know, there's, yeah, for people who want like absolute certainty that their data is not going anywhere like that's that's the solution i was talking about it with a friend earlier like companies will have just big you know like sort of servers mainframes or whatever to run all of this why would you spend all your tokens with anthropic when you know you can run it almost for free on your own machines yeah and then there's also um all of these third-party providers and so if you go to this link which i'm sharing in the chat artificial analysis for any model you want to run you can look up and and you know see who offers um basically servers to run that model for you and you know you can partition them private that sort of thing so yeah that's another solution if you can afford it also i haven't tried it yet but open router have a cool feature called fusion which i think tries to pick the best model for a task and sometimes combines two or three to get you like a sort of um panel of panel of answers um yeah definitely moving towards local and also like getting two or three agents to kind of work on a task together before coming back to you uh got one or two more here before we wrap up so um if you if you have time do you have time yeah i can stick around yeah cool all right so it's uh it's quarter to eight in the evening for me so i've i'm i'm done for the day yeah that's great um so uh abur abarian asked um where do you start once you've got claude linked to google ads and meta are you using Claude for any like ad buys these days as a solo business entrepreneur no uh so far so far I'm lucky enough not to be not to be buying ads um I've heard so okay so meta released a CLI for their ads platform like a couple months ago and like giving Claude code access to that is apparently quite good uh windsor.ai is a nice MCP for like plugging everything in so you connect you know Shopify, Meta, Google to Windsor, and then you just have one connection to Claude, and that handles, like, all of it for you.

1:44:15I've heard good things about that. But no, I haven't done much sort of, like, paid spend in a while. I just interviewed a guy who's running, like, a growth agency this morning that's become fully AI native in the last, like, eight months. And they pipe everything into BigQuery, and they've built a bunch of, like, Python scripts that any agent can run to kind of, like, you know pull the latest information and then they have scheduled tasks that check whether ads are fatiguing and so they'll kind of get a ping in slack that says hey like this campaign looks like it's starting to fatigue uh maybe we should like improve it so loads you can do but not much i've done personally in the last year or so that's fair that's fair um gwpi card said how do you set up background agents either running periodically like a cron or on a trigger like when X, launch Y, automation workflows.

1:45:09You kind of covered this, but do you have a direct answer to that? Yeah, let me, there's actually, that's a good, there's a good part of that question, which was like, when this happens, do this. And I didn't cover that. So let's cover that. So, you know, my whole session, I was in Claude co-work, but there's also the code tab, which is honestly very similar to working with co-work, just more powerful. And if we go into routines and then new routine, I think it's local. You can basically, nope, it's not local, it's cloud. You can trigger it on an API. So you can basically say to Claude, when I send you a webhook, when I send anything to this URL, I want you to then go and run whatever task I have using my connected MCPs.

1:46:07I actually, I did have a slide on this and the one thing with, it's disappeared, here you go. The one thing with cloud tasks in code is that they require you to have a GitHub repository. So you can run a local routine on your computer without GitHub, but in order for Claude to basically sort of like run code inside of your folder in the cloud, it requires you to have a GitHub repository. There was a question in the chat about like... Evan just asked about GitHub. Yeah. Yeah. It's just a way of like... Just so you know, we're going to do a whole stream on GitHub in a couple of weeks. We're talking to someone at Microsoft who can come on and give it like the whole rundown.

1:46:51But yeah, so you can give the Sparknotes. Sparknotes is like, it's a way of collaborating on code with other people. you basically host your whole code base in the cloud and it's usually a private repository but people also give away code open source for free using GitHub and when you have your folder connected to GitHub an agent can basically copy everything over run code in the cloud on a different server and then sort of like shut it down very easily yeah and like we explained earlier with plugins github there's a really good github plugin for claude code so you can literally just set it up give it access to either all your repos or select repos and you can say like hey let's you know use the github plugin and let's push this code yeah i yeah i did computer science at uni but in the last year i haven't run a single git command myself i just yeah you just know i just ask claude i'm like yeah get onto a new branch and like submit the pr when you're done and run it through the tests but like yeah that's what i told the github people i was like if you're coming on this is what this is the way we're talking about using it because this is how i use it yep yep yep yep the one thing i do actually one thing i started doing a few months ago with git is using work trees have you used work trees before uh claude does it automatically so yeah it's super helpful yeah go ahead so basically the journey that i went through and that loads people go through is like you discover claude code and you have one tab going and you're like hang on a minute and then you start getting like four agents going at once and you feel like tony stark with jarvis but then the agents start colliding and they start like submitting prs that kind of conflict with each other and everyone has that story where they're like oh four hours of work just got deleted because some agent thought it was irrelevant um so work trees basically clone the folder on your machines so that everyone's got their isolated folder and they can like the agents can collaborate without overwriting each other's work yeah so ryan karkson i don't know if you're familiar with him he's yeah i've learned so much from just his interviews genuinely same 100 and uh he was saying that everyone needs to update their thinking from coding locally to coding in the cloud uh because of that reason where basically if you have too many work trees on your computer it's gonna fill up like i personally had this problem where i was trying to download a model and then it said there's like no space on your computer i had this yesterday i I had like nine work trees and Claude was like, well, each one is like several hundred megabytes for some reason.

1:49:28So the tip for that is tell it like when you do a work tree, send it to the cloud. Because you can now have Claude send like different agents to the cloud, even though you're working locally. You can just have it like do that for you. And I think that's the tip there. Nice. Let's see. How transferable is agent building from platform to platform? um if we review on one today then i use something different is it the same process idea great question great question so a lot of people they've been talking to chat gpt or claude chat for like a year right and like they kind of feel like claude knows them or chat gpt knows them but if that account got shut down they would lose everything and actually i saw a message earlier asking like why would i move from chat to co-work so if we stop relying on like the memory of the llm and we start relying on context whether it's files on our computer or whether it's like a notion folder somewhere we can actually just drop in any agent we like so you know i just demonstrated earlier opening like kimmy glm and claude in the same folder they all see the same instructions the same context if i decide tomorrow you know if anthropic goes bust tomorrow and they shut down claude and i have to go and use codex i just pick up where i left off um so um which i really hope doesn't happen by the way that's not a prediction um so um yeah like once you move away from the basic like cloud chat products to code work or code or codex you're actually making life like more portable like it's easier to transfer between them yeah yeah and everything we've talked about with maybe the exception of plugins um like is transferable so if you have your second brain on your computer that's transferable to any system you use um if you have skills those are downloadable files that you can take into.

1:51:20I think I said in the chat earlier to answer someone's question, but I think Grok now uses skills, ChatDBT definitely uses skills, so it's transferable to those. Gemini, I don't know where they're at on skills. They might get them next year, yeah. Yeah, if you're lucky, fingers crossed. And what was the other thing? The only thing is scheduled tasks, but you can copy the logic from it. Well, yeah, it gives you a big instruction as long as you've got the right MCPs connected. You mentioned ChatGPT for work earlier. I tried it a couple months ago when they had these workspace agents. They're really easy to set up.

1:51:56I think they've actually made the UI a bit easier than Claude for scheduled. So my problem is they're too easy to set up and I was spending 800 million credits per week or something like that. Obviously, this is on the subscription, so it's not me actually paying this. but they do this thing called like credits which is kind of confusing and you run out pretty quickly and I think it's BS I think they need to fix this ASAP because it makes it impossible to use when it is a very useful tool so your mileage may vary with workspace agents I was able to automate a ton of stuff following basically the same workflow that you set up today but in chat to But the problem is, you know, I hit the limit on workspace credits and like a weekend and there's three weeks left in the month.

1:52:46So it's not a perfect system yet. But that's why I think it's important that we push for local models so that some of this stuff can be done locally. And if you use an independent agent like Hermes or OpenClaw, some of the models like for scheduled tasks can be run locally on your computer, even with OpenRouter. Right. Well, this is actually, yeah, this is something, if we have like two minutes, I want to talk about my new favorite agent harness. Let's do it. So do you use Vercel? Have you come across Vercel? Yes, we've talked about Vercel before. I use it occasionally. I try to self-host a lot of stuff, but yeah, I like Vercel.

1:53:26So yeah, it's amazing for just deploying any code you write to the internet. Cloud Code works with it really well, makes it so simple. they've released this framework called Eve maybe like a few weeks ago and it's a framework for building agents and basically you like open this up in any folder on your computer and you just like it needs code code so you give it this command right like start a project and and it it basically creates an agent in the cloud that you can sort of reach across any channel so like they have got channels for where is it now telegram it's not here they got like telegram slack um um whatsapp you know you can run it as like an agent that lives on your website that people can talk to um and last night i thought okay well i have this folder on my computer that i'm using for my like health and fitness so i've got my metrics coming in from garmin from training peaks you know my blood work i open cloud code in this folder and i can have like really interesting conversations about like supplements i might want to take or like my resting heart rate trend you know all this stuff and like i've built a whole health dashboard that i could show you another time so i ran eve i ran eve inside it and within half an hour i had an agent on telegram called health bot and now at 7 a.m it sends me a briefing at 12 p.m it looks at my my fitness pal and sort of says hey you've had a bit too much sugar today like maybe tone it down in the afternoon um you're coming back and teaching me how to make this i literally need this it's on my goals it's it's it's yeah it's because i kept thinking like if something could just ping me throughout the day like that would that would actually change my behavior and so and i also thought like i want to track my supplements but i i can never find an app that makes it easy And so now I was talking to this agent in the gym this morning.

1:55:22And like, you know, I was like, here's my weight. And I've just taken these supplements. And then it pinged me today at like 130. And it said, looks like you've only had 60 calories so far. And you've got a ride coming up this afternoon. So you should probably try to, you know, it was amazing. And like on Telegram could be on WhatsApp. And it was just with Eve. And it's all running in the cloud. And it can switch any models. So it's on Sonnet right now. But I'll probably pull it down to something cheaper like GLM. Amazing. So yeah, I think this is like, I think this could be like the open claw killer.

1:55:52I think this is like so easy to use if you just point clawed code at this website. Yeah. Yeah, I think we're still waiting for that thing, that application that comes out that abstracts away the like pointing it here and pointing it there and just like works out of the box. Like, because I think I find the hardest thing with Hermes or open claw or, you know, perhaps even Eve, although you tell me how easy it is, is that, you know, you have to sign up for these like endpoints and and you know make sure that you have like a clear gateway and that's a little bit intimidating if this is your first time doing this you know what i mean yeah for telegram they made it really easy for slack um it looks pretty easy because like we've all been through that process of like getting a token and like provisioning a new app just so that your agent can sort of like send its first message and um yeah i agree i think like we're waiting for a platform that just kind of allows you to spin up an agent with the right context and the right channels i think eve are like pretty close to it and i'm i'm i'm bullish like after that experience of like you know 30 minutes to like a fully connected health agent that reads and writes and uses tools in the cloud i don't even know where it is like it's in a versell server somewhere right like i didn't have to like i didn't have to spin up a vps i didn't need to get my mac mini running my computer doesn't need to be on it's just like it's just on it's amazing okay well have to have you back to talk more about that.

1:57:19Someone asked before we wrap, what's the typical monthly cost of deploying an agent like the health bot? So in the last 24 hours, it seems to have cost me$2. And that's on Sonic 5. And that's probably like 40, 50 testing messages. So... What if you switch that to DeepSeek before? Exactly. I'm going to. And then I think it will be like, it will cost me pennies to run. And if it costs me like, you know,$10 a month, but I actually changed my behavior and like have this, you know, yeah, definitely worth it. Yeah. Well, James, this has been amazing. Thank you so much for spending the whole hours with us.

1:57:57I think everybody shout out in the chat, you know, if you've found this for the questions. Yeah. Thank you everyone. And if you didn't, if you didn't get your question answered. So what I do after we do all these live streams is I go and I take the transcript and I turn it into an article where I try to answer everything and put it in a nice order so it's easy to follow so definitely stand stand by for that i'll work up work on that for everybody but uh james where can people go if they want to learn more about you and your work uh you can go to agent accelerator.ai where i like run a live program teaching this stuff and there's also an online course you can take uh i post loads of stuff like this on linkedin so you can just find me james mccauley on linkedin and i share all this good stuff uh and then i am doing more on YouTube.

1:58:41So subscribe on YouTube and you'll find out probably tomorrow how to run Kimmy and GLM inside of Cloud Code. Oh yeah, send that. You definitely better email that to me because I'm very curious. I will. I will. Alright everybody, thanks again. Thanks everyone. James, hopefully we might work together again on some projects. I think that's going to be a good work. I've heard there's some stuff cooking. I think you might see me again on the Neuron. Very cool. Very cool. Well, for people who don't know, we have a Neuron Academy our own course platform that we're building right now. I'll share a link to that.

1:59:14If you like this, we create more content like this on there. You know, we can only do this once a week, but loads of loads of material on there for everybody. So with that, thanks for joining in and farewell for now, humans.