BONUS: AI Skills vs Agents vs GPTs: Which One Do I Use?

32:23So you could say to coworker, hey, use my AI skill of the day creator skill and write me an AI skill of the day to learn all about skills. And because you've already taught it how to write in the format of the neurons AI skill of the day, it knows how to then convert that request into, you know, it'll go research the web. It'll go make sure it understands the assignment. It'll read the docs related to that, you know, what skills are. and then it will come back and it will write it in the exact format. So co-work is like the agent. It reads the skill file. It learns, oh, this is how this works.

32:59And then it goes and it takes the action and it does it according to what the skill file does. You can kind of think of skills as like it's a prompt and then a bunch of other stuff that teaches it how to use it. So what you want to be doing is you want to be working with co-work to create your own skills so that you could say like, hey, make me a skill to do this process. Make me a skill to do that process. So you don't have to memorize prompts. You don't have to be constantly prompting it and trying to get it to work. You can just literally say like, add this skill or use the XYZ skill and it'll already know what you want.

33:30It's really powerful. It's probably the biggest unlock of how to use AI, I would say right now. Yeah, and very much, you know, use the tool you love. I just saw the comment here I wanted to address real quick that you've been using ChatGPT exclusively. Is there a better AI company you would recommend for the group insurance service business? Me, no. The truth is, if you've been working with one for a long time and you're good at working with that one, like you feel like you're able to get the results you want, I wouldn't mix it up. You're looking at small-scale changes from one to the next and many percentage, tiny percentages quite often, with the exception of like this week, and there are other weeks where that's the case too.

34:11But it always flip-flops. You know, you're never going to always have the best one. So, you know, the AI you should use is the AI that helps you do what you need to the easiest, the best, and that you understand how to communicate with. So if you're working with ChatGPT and it works great for you, cool. Use ChatGPT. Traditionally, that's kind of been my go-to. I still use Claude some. I use Gemini a little. I use Metis AI in my glasses that I, for some reason, can't have. I want to give a challenge to what you just said. Go ahead. What is your challenge? If you've been using ChatGPT exclusively, I would challenge you to try Codex because Codex is much more powerful than the ChatGPT app.

34:57And eventually, according to rumors slash leaks in the media, they will eventually merge them into one application. Supposedly. Codex is ChatGPT on the juice. Yeah, it's you can use it with you can use it with your current plan. that you're paying for, you just have to download another app that's separate from JAT-GPT. I would try that because that's going to be the thing that's going to give you the most

35:27acceleration from where you are now. It's a real bummer right now that Anthropic and OpenAI both shoved code in the name of those tools because what I'll say is that that word is there intimidates a lot of people who are not engineers or developers, please know those tools absolutely 100 % have value to anyone. They're very much built for work and setting up automation. Do you have codecs open right now? If so, show them how to toggle between the work and the code version because that's really helpful.

36:05Dun, dun, dun. One moment, please. Mm-hmm. not a problem yeah so basically what what cory's saying is that like it seems intimidating because it's like codex or claude code which is why claude uh eventually made co-work which a lot of people are familiar with now but you can actually switch um codex to be more focused towards like knowledge work essentially so that it doesn't speak to you in this kind of abstract mystical characters that only engineers understand but it kind of talks to you like co-work does so grant my codex on this work machine the the thing's not there oh interesting well i can share mine oh wait never mind i found it i found it okay once let me pull down the screen i was going to use to show a skill here and switch not a script and then we'll and then we'll go back to we'll get back on topic after that, I promise.

37:07Yeah. Yeah, because I want to try our best to stay on topic for folks who find it hard to learn from, you know, us. We're resources, so ask for what you want to know and we'll focus on it. Absolutely. This is Codex. If you've never used it, don't be that intimidated. You come in here and you'll write prompts just like you would, you know, anywhere else. What's the difference? You'll spell it better, hopefully. than a bluebird and a robin. Friendly reminder that typos don't really matter. You don't have to ask it coding questions. It will absolutely do anything else you want, just like you do in ChatGPT, except you get access to automations, sites, which is a cool new feature, plugins, skills, a lot of things that honestly just work better.

37:57It's working on your computer. I just love it. The truth is, as a work tool, it's great. So is Cloud Code. They're both great. We're really in a cool moment for such things. But I'm probably in work mode. But what you should know is if you go to settings, first thing you're going to see in general is, do you want to code or do everyday work? So the first time you set this up, it will absolutely bring this up on a screen right at the beginning. And it'll be like, hey, do you want more technical responses and control or do you want the same power but less technical detail? You switch to everyday work, back out, and there will be a number of things that it won't do anymore.

38:43It's barely visible from just glancing. Yeah. You know, one of the things I like that ChatGPT doesn't do is the ability to go between all the different reading levels to go into fast mode, which is ridiculous. That's the one running on those Cerebrus chips. If you don't know what that means, look it up. They're ridiculous. It's also ridiculously expensive, just to be clear. So you don't use fast motors, are you saying? So you can pick between different models. You can choose your reasoning level. I usually live on medium, but here lately I upgraded to the pro plan, so now I live on high because I like to live a little and who knows.

39:21But either way, they both work great for me. but now I want to take just a moment and go back over to the skill that I was going to show you while we were talking about skills Codex is connected to ChatGPT but you can't run it online I don't believe unless that changed when they gave us mobile use you can kind of connect them yeah I think you can I built an app last night that wasn't working right on my desktop computer and it was working fine. And then this morning I got up and the report it was supposed to send me came in with an error. It didn't work. So I opened Codex at the coffee shop on my phone while I was waiting for my latte and said, hey, this didn't work.

40:07Fix that. And it opens up my computer at home, fixes it, and by the time I'm home, it's already at my work computer ready for me. Really handy. Pretty awesome. Okay. so that's that to the other screen okay while you're pulling that up i want to address root castro silva's comment wouldn't it be sensible to also plan for portability instead of being dependent on one provider yes which is why we recommend that you focus most of your attention on building reusable skills so the thing that we're about to show you how to do because both platforms let you use that and now grok lets you use that so the idea is that there are files that you can download you can take with you to whatever platform you go to so yes it's helpful to have a project folder with all of your um you know context in it but of course if you have a folder on your computer where all that context lives you you know it's the same thing uh and you could take that with you anywhere obviously most of them also let you download your uh your data your archive yeah yeah but but but skills is like the same concept where it's like it's a it's a pre-packaged workflow that you can take with you whichever one you're using so for example uh for the last couple months i was using claude exclusively for all of my work-related stuff and then i switched over to chat tbt to test it out and i just brought all my skills with me and it works pretty much exactly the same way as I had before.

41:40Yeah. We've reached a point where they're all good. With that said, just like in anything else, each has little nuances, things it's maybe a little better at, a little better suited for. The differences are generally much more minimal than people in the AI space like to admit. Usually they're close, you know. And depending on your job, you may never notice the difference. Like if you're an electrical engineer, you'll notice a difference. But if, you know, if you're a bookkeeper, those are tasks it's been really good at for quite some time. Probably you're, and I don't mean to diminish bookkeepers, please know that.

42:18I just mean it's a task it's been quite good at for a few model series now. And you're going to notice less of a difference there upgrading than you will in others. Kind of the way that happens is different skills just take it longer to figure out depending on its training data. And kind of once it hits the, I'm very usable here, you're going to start seeing smaller and smaller returns with new models. So talk to me, what are we about to learn? I'll tell you what we're about to learn, my friend, we're about to make a skill. Let's do it. Sorry. I love it. No, no, no, I love it. And so for context, I have started making some shorts each day for our YouTube channel.

42:59and in doing so I was like I need something to spice these up a little and I decided you know what I want like sketched out b-roll type images I can bring up on screen to illustrate things periodically that just you know that are like you know they flick in and flick out after three seconds and I've kind of settled on this style I've been tossing them in I have that one the other day this is one about agent loops cool i did a you know what's the difference in a prompt and a loop short the other day if you wonder that question we got a short for that now uh but i've got it kind of dialed in how i want it like it's essentially the same kid it's the same style of sketching so once you have a thing that you do regularly and it does it pretty much how you want and you're like you know what i want to be able to do this anytime You can come right down here to your chat box and say something like, create a skill that will generate images of this style and orientation.

44:12when I request a...

44:21We're going to call it a B-roll sketch. Now, the way that he's able to do this is that there is a built-in to Chat2BT business account, I believe, and Codex is a skill creator skill, which knows exactly how to create skills and can give you a one-click executable file. If you try the same exact prompt and it doesn't give that to you, that's because you either don't have that skill installed, which it should come in natively if you're on a business account. If you have skills, you have the skill. Yeah, and this works in Claude as well. So I saw a ton of people using Claude in the chat. This will work for Claude too.

44:58And this is literally all I do. I'll hit enter. It'll take it a couple of minutes probably because essentially what it's doing is it's building out a couple of files that knows what it wants. It's deciding how to describe it so that ChatGPT can replicate it consistently. And so it's going to package this as a B-roll sketch skill with the trigger style rules, vertical format, and prompt template baked in so future requests can produce images just like the hand-drawn examples without restating the full direction. So it's going to do that for a minute or two. And then it's going to come back, and once it's ready, it's going to give us the option to try it in a chat and the option to install and or download, which is really easy.

45:39You'll hit the install button. It's there. I can hit download and then go over to my codex and literally just add a file to a chat and say, hey, install the skill. Yep. Done. Yeah, it's very simple. They've made it so user-friendly. Both OpenAI and here it is. That's what it looks like. It has that kind of multicolor cube animation. So if you don't see this, that means it's not working. and it's gaslighting you because you don't have access to it. So try it in codex. It's installed. Yeah. And then it's, okay, it's installed. Now we're going to hit try and chat just so you can see how easy this works.

46:14So now that we have it, I'm going to delete all of this because it's going to give us an example. I'm going to say. You're in another window that we're not seeing. If you could switch. My apologies. Okay. So over here, it gave us a chat to try it in. Thank you, Grant. Could you zoom in a little? a little bit. One more? It looked really, really big. That looks good to us. Okay. All right. And what you're going to see is it'll show up down here as B-Roll Sketch. You don't always, it won't always. If it's not there, you can just tell it, use your B-Roll Sketch skill to do X and it will load it, run it, and it will work.

46:56Basically, if you use that language, it'll know to check its skill registry and see what's available. uh and are they in here no yeah they're they are i couldn't find it oh yeah so if you hear yeah oh wait yeah if you hit plus um and if you um scroll and then to more more um they'll be in there so like for instance that was like a beehives connector oh it's maybe it's the connectors that show up there yeah those are my uh my agents okay never mind disregard they're in the skill registry which you can find from settings. I'll show you that as soon as we finish showing one quick example. So we're going to say, use your b-roll skill.

47:43I'm not even putting the dashes in there. It knows what those are. It knows what it means. It knows what I mean when I say that because it has good memory. Use your b-roll skill to create a sketch

48:01showing the difference in a skill. Ooh, more typos. Typos for days, man. Typos for days. A project and a custom GPT. So, God knows what, I don't know how it's going to illustrate that. It's probably going to have more words than I want. But what we're going to see here is, I've got to bring this up, or I'm not going to be able to see the image. on the screen.

48:30There you go. Yeah, that's fine. Okay, yeah. Here we go. Yeah, because the image will come up full screen as soon as it's done. And so it's making the thing, which obviously takes just a minute. But the catch being that I've given it instructions that it already knows and that it understands how to use. It understands what I want just by me saying the words use your B-roll sketch skill. Like it knows, ah, got it. He wants one of those things and moves on. there so instead of me having to go in every time and say i want you to make a 9 by 16 formatted sketch on an off-white background that looks like a drawing uses a little boy with spiky black hair and uh use that to illustrate this instead of that i just go and i'll say the thing okay see this made more of an infographic honestly uh now here's here's something i've been wanting to really here's something i've been wanting to bring up related to this so i was trying to create the equivalent of that google gem that i showed earlier in chat gpt and i wanted it to use reference photos that i provided in the chat well because one of the things that these um ai agents are not good at is manipulating images and moving them around and putting them places for some reason they can't do that really um at least you know if you know how to get them to do that please shout out in the chat um what what you have to do is you have to actually go to your desktop open the skill file and put uh create the folder with the reference images uh that you want you have to create that folder and then manually put the reference images in or maybe you could ask codex to move the files from your computer i think you can do it in codex but you can't do it in the mobile app right now, or sorry, the web app right now.

50:19So that's the way that you do that is what you would do is you would take that reference image that you really liked, you would have codecs add that to the file, the skill file in an image references. Yeah, format. Yeah, and do it that way. Exactly. So here what we're looking at, actually, I'm going to download this image and figure out a way to share this with you because it's actually a pretty helpful resource for what we're doing that might be worth saving. It's going to kind of break down. It's an infographic breaking down exactly what we're here to talk about today. So I'm going to stop sharing.

50:54Oh, I was going to show you where skills are. That's the other thing. Yeah, please show that real fast. If you have a business account or if you're in a product, you'll be able to see it. Yeah. You go down here to your settings and there's a skills option. If you don't see that, your chat GPT doesn't have skills yet. But it will in Codex if you go load it there. And you come over here and you're going to see all of the skills I've installed. Some of these are mine. Some of these are not. But they're all ones that are installed in my ChatGPT instance. Here are ones I've created for a variety of tasks in this specific tool.

51:35Like I have more in Codex. and here are any inside our company that I can go and pull from that other people in the company have created. So like, you know, sometimes the best answer for a skill is to go see, has somebody else made that yet? Because if they have, wonderful. But that's it. And you can go right here and create a new skill, but my preferred way to do it is to do it inside a chat where I try to get the result I want first, And once the result is zeroed in the way I want it to look or to read or whatever, then I say, okay, that's the one. Make that a skill. And I've had a lot of luck with getting those to perform really well for me.

52:20I think a good rule of thumb, and I think, Corey, you talked about this when you taught our company how to create skills, is if you do it more than twice, it's worth a skill. So if there's a recurring task that you do more than twice, turn it into a skill. Please save yourself the prompting headaches and the memorization. I might need this again someday. Yeah. And you don't want to because like it's a hassle. If three months from now you need to do the same thing. You know, you saw what what the chat history looks like. You're never going to find the conversation where you did that. Its name might be nonsense.

52:55You might have typed it in a conversation. Not about that at all. Yeah, exactly. So let me show you another example. The free plan of Claude. on my work computer. Oh, okay. I could show, I have my other computer queued up to join us at some point where I can show what a non-free plan looks like. But in this case, so for all the Claude users out there, the way that you find skills is you click over here on the side panel, you go to customize and you can go to create new skills right here if you want to create one. Or I believe you go right here to this little scroll icon. Let me follow my own advice and zoom in.

53:35So you go to the scroll icon and then you'll see all of your personal skills, which you might not have any if this is the first time you've come here. Word cut cut. But then, yeah, you can see all of my attempts to play around with this. And then you can go to either browse skills and you can see all of the skills that Anthropic has created, which in this case you can see they have create beautiful visual art, build an MCP, create a theme factory, et cetera, et cetera. um i scrolled over that part in open ais too i forgot to mention that they do have skills you can just download yeah but you know to create and there's three different ways to create it you can create a clod you can write skill instructions or you can upload a skill that you've let's say like created in codex you want to upload it directly here you can do that so yep um somebody had a question in the chat how is a skill different from an agent um think of the skill as like your prompt, your project structure, your instructions, your intent that you want to give the agent.

54:38So if it's for a recurring task that you want the agent to do for you multiple times, the skill is all of that in a neat little package that you give the agent. So it's sort of like knows you. It's like, ah, I know. I know how this works. Like, you don't have to tell me because we've done this before. And it's like solved. and another thing to keep in mind with skills is that you can version them so you can continuously update them over time in order to get them to uh you know in order to improve them if you catch the ai making mistakes you can add like hey make sure that you always do this or you never do that and you can uh course correct it and one good thing to keep in mind is that you should have it label itself so like this is like skill creator 1.2 skill creator 1.3 so that if you do um create multiple versions you can kind of keep track of it and then you can keep track of that with your team so that you and your team are using the latest version of each skill that's a really good call out grant thanks um so let's see you got some questions in the chat can you show a writing a skill for an interesting to read facebook response i actually can um let's see let's see here um why don't i do one here uh the google universe to my knowledge does not yet have skills is that correct grant they have it with gemma which is our open source model but that's probably too technical for anyone on this stream i don't let me check my personal gemini account and see if they have it yet yeah me too i'm i'm looking i don't think they're really sleeping on it and i to be honest i don't know why they won't for long though this is absolutely a kind of a standard thing that's it's being widely and broadly adopted now like it's it's used in all of the coding agents we didn't explain what an agent was compared to a skill i was looking for my notes i have a really good explanation of that somewhere we kind of just did but that's okay we'll we'll break it down even simpler an agent requires a number of elements that a skill does not let me say that an agent is a combination of a trigger meaning something that makes that thing go run some amount of tool use some amount of maybe verification other things and it's bringing all of that together to perform a task a skill is just saving a specific workflow so like one One of the differences is like skills don't have, I mean, I guess a skill has a trigger, but the trigger is just you saying use the skill.

57:24And so one of those, so for example, if I want to do something hands on, I'm probably using a skill to help me through it. If I want something to happen like magic and just appear in my inbox in the morning, for example, I'm using an agent. I have an agent that I built last night that is running a small little Python desktop app that Codex made. And it does a lot of YouTube research around things that are interesting right now, making sure we're not missing things we should be talking about, teaching about, and learning about, for that matter. And I set it up to where it's going to send me an email every morning.

58:02So every day at 5.30 a.m. Central Time, it wakes up. It calls the YouTube API. It's going to go out and it's going to gather all this information from all these specific channels I've given it and other things. It's going to come back. It's going to take that data, analyze it, make a report. So we're calling like multiple tools here. And then it's going to take that report, save it as a PDF, drop it in an email, go to Gmail, email it to me, to Grant, and to whoever else. So like an agent is definitely a more complex workflow tool. There are a number of things you could do the same, but you can't use time triggers.

58:43Schedule triggers, for example, if that makes sense. I hope that's helpful. If that's not a good enough answer for you, let me know and we'll come up with a better one. I love my gem to create lyrics for education songs, but creating video clips just seem to... Yeah, video's high. Takes all to compute. It takes a lot of compute. It costs a lot of money. It's really cool. It's doing it a whole lot better. yes you can get a copy of the chart i made i'm trying to figure out if i can drop that into a youtube comment right now oh okay um because that was i was like oh that's actually a really good handout for this class we're teaching can you can you share did you uh oh yeah are you're talking about the image that you created yeah the image i created with the skill a minute ago yeah so what's another what's how do i want to add on to what we're talking about um we have a whole stream by the way agents for beginners uh i think i want to share that in the chat if you haven't watched it yet um because we go much more in depth into agents and what what they're made up of and i think that'd be helpful if you really really want to understand what the heck they are and so i'm going to go ahead and share that in the chat so um i would just say like the the simplest way to think of agents yeah agents can use skills but skills cannot use agents well depends on if you're using sub-agents which we'll get to in a minute yeah but uh technically as complex as you'd like it to yeah technically that's kind of true unless you're doing like some pretty intense agentic engineering um but yeah an agent you can think of it as an ai that can take actions towards a goal so you give it a goal you give it instructions you give it access to tools and then you let it make decisions on your behalf it can browse it can click it can read files it It can run code.

1:00:32It can use tools. It can work across multiple steps until it's complete with the goal. So like research a topic and build a spreadsheet or review these files and then draft updates from a source doc or fix this coding issue, like get this PR and then go like ship it or give it to me to review it before you commit it and push it, you know, things like that. So it's like an AI that you give it a goal and it goes out and it does things for you. And in order to enable it to do that, you have to have what's called a harness, which cowork or codex, those are like agent harnesses for an agent. And we go into all of that in the agent video that I just shared in the chat.

1:01:15So you can learn a lot more about that there. By the way, there's not a way to share an image in our YouTube chat. I'm telling you, images, I think for security reasons, because images can be very, you know, have malicious code. I wonder if I can screen grab a screenshot and paste it in.

1:01:39If not, Grant can include it in the newsletter to download. I'll do that. I think that's a better way. I'll make sure Grant has it so it can go in. Today's Thursday. Tomorrow's newsletter? Yes, Friday. Okay. Yeah. Okay. I'll make sure he has it. Sorry about that, everyone. I really thought that would be an easy thing. Yeah. It's okay. We can do it. We'll include it in the post that goes with this as well. So check this out. So somebody asked, can I create a skill to make a Facebook response, a funny Facebook response or entertaining Facebook responder? And so I said, let's create a skill together using your skill creator skill that can make interesting to read Facebook response.

1:02:24To me, that is a direct comment, single sentence, A funny snarky response includes short phrases like LOL or WTF. Asks a question at the end, starts with a ridiculous claim and then explains it. Makes a trendy pop culture reference related to the topic, using web search to look up something related to the comment I'm responding to, and talks like crack Gen Z brain. So to me, that's what would make an entertaining comment. Here's the example output, right? Someone said, I don't understand why everyone is obsessed with Stanley Cups. They're just water models. and then the output is okay but sailing cups are just water bottles you're literally right statistically speaking paying 45 for a cup it's just a personality disorder with a warranty not gonna lie the whole thing is giving hydroflask 2.0 but make it louder and people literally stampeded a target this is doing too much for me so i'm gonna say these examples write too much yeah they need to be more punchy

1:03:32We got a good question here. I will answer in a few, whether Codex is an agent or a harness, if you'll hang in there. I want to let Grant finish this before we move on. Good note. The skill is trying too hard and doing too much.

1:03:47I want to say also include typos.

1:03:57All right.

1:04:01okay so it's going to take a minute to go through that what's the what's the question while that's going i'll tell you what i've got an answer for this too uh the question is is codex an agent or a harness and the short answer is yeah codex codex is a product the the product codex the tool you download is an agent uh it's It's a coding agent for software development. The Codex Harness is the implementation layer that makes the agent work. So maybe that's a little bit splitting hairs. No, no, no. That's good. I'll allow it. Okay. Okay. Like I said, it is splitting hairs a little bit, but I wanted to make sure we had the right answer.

1:04:46And that's according to Codex because I was like, that's a harder question than it seems like on the surface. I absolutely cheated because I'd rather have the right answer than make something up because that doesn't help any of us, does it? No, that's the right way to frame it. Yeah. Know your limitations. Recognize when you're like, I don't know what the hell I'm talking about here. I won't say that. But I just kept thinking both. Could be either maybe. And that makes a lot of sense. the core agent loop and execution logic under the codex experience. Yeah, I think that's right. Yeah. So here we go.

1:05:28I just have closed the loop, speaking of, on this. So now what it did, it took my note, and it's turned it into a skill file here. This is what this looks like on Claude. I can click Save Skill, and then it saves it to my skills. I can click Manage. I can go look at it there, the screen I was showing you before. So I'm going to go back. then I can also download it or I can click into it and I can look at the the language and you know a lot of times maybe it writes too much or it over cracks or it over steers so you can actually you know download this come in and edit it um you might be able to edit it here Claude is wordy yeah you can go in and you can edit it yourself and then you can re-upload it um you can like I said you can download it and then all of a sudden you can now edit it yourself but yeah that just that's how you create one and you know you could just keep iterating on it you know when when you're in the middle of a chat you can say hey i want to update my facebook response writer um skill it's saying this too much or it's doing this trend that's annoying to me or you know let's make it vary its ideas more frequently things like that and and you know continuously update it that way yeah we got a good question about a uh an analogy here and i'm in automotive terms a skill is filling the gas tank an agent is removing an engine good analogy i i feel like maybe a better analogy would be a skill is the gas tank because it's there it it knows what you want you want gas it knows how much gas you want you want 16 gallons uh an agent is grabbing the pump and filling it up i i feel like no i you had me on the first one i would say the agent is the mechanic the agent is the mechanic and you say hey go fill the gas tank um and the mechanic decides how to do that the mechanic has all of the tools in the auto shop well so yeah if you're pumping gas we're talking gas pumping you're the agent you're the one going and getting the thing and pumping the gas.

1:07:33The agent is... Yeah. I think that's the best way to say it. I mean, the way I would say it is you're kind of like the customer at the auto shop, and you're saying, hey, I would like my gas... The gas tank is maybe... Someone's filling your gas tank for you. Yeah, exactly. I would say oil change is maybe more accurate because, at least in the state I live in, you can pump your own gas. Yeah, but you can't... I guess you could change your own oil you can here you can change your own oil to here right now i'm just lazy and i don't do that but yes you can you can absolutely but it's more common practice for you to go and have someone else do it uh yeah it's more socially acceptable an agent would be like you made an appointment with that person to do it tomorrow at two yeah and that person's going to change your oil tomorrow at two while you sit back and watch the television in the lobby this metaphor may have gone away from us but we've gotten the right the right way to think about it is the agent is the mechanic who you are asking to go fill something up for you and that mechanic went to school um it's been given a list of instructions for how it it does each action um it has a list of actions that it can take it has all the tools available to the mechanic in the auto shop and it can now go and decide the best way to take that action for you so you're the customer the agent is the mechanic the skill is the like i guess like the tool belts the ways in which it knows how to do things yeah exactly exactly all right so let's talk about plugins old plugins so now that now that we've made skills good and complicated i i think skills are a lot simpler than people think imagine it as like a recurring task that you can do uh in any chat window and you can just add hey run this run this task and then you can give your agent access to those skills so that it can do it for you.

1:09:28Now, a plugin is, well, before we talk about plugins, we should probably talk about connectors. Before we talk about plugins, we should probably talk about connectors really fast. But to close the loop on plugins, a plugin packages multiple pieces together, like skills, potentially different agents, connectors, hooks and tools, which are just like functions that the agent can call, all into one package so that you can give that to your other teammates. So instead of saying, hey, teammate, download these 30 skills that I've created, you can package all of them together in a plugin and you can make it available to your team.

1:10:03Yes, if skills are the tools, plugin is the toolbox. That's good. Before we get too deep into plugins, I'd like to take just a minute and see, does anyone have any questions on what we've talked about so far? That way we can make sure that if there's anything that's not clear, we can try our best to help you clear it up now. That way you're still able to follow as we move through the second hour here into grants, a slightly more complex deep end. None of this is that difficult. You can understand every bit of this. We didn't at one point either, and by golly, we learned it. So you can too. I find what often happens for me is Anthropic, and it's usually Anthropic is innovating here, will come up with something and they'll say like, okay, we created skills.

1:10:50And then my ears perk up and I say, oh, that's cool. What is that? And then I don't really get it at first. And I'm like, okay, I guess this is useful for engineers. And then, you know, a month or two goes by and then it snaps into place. And I'm like, oh, I get it now. And then all of a sudden I switch all my workflows in that direction. And it's like, I get it. Like, it's funny in the process, you know. can you transfer projects from ChatGPT to Codex? Corey, what do you think about that? I don't think so. I mean, you could manually create a project in Codex. Like, essentially, there are, they're used a little more like folders in Codex, but there is a project thing, and you could do that.

1:11:42I mean, essentially, what you do in Codex is it's a little less necessary because you can go in and be like, all of this stuff is in this folder on the desktop. Just give it the path to the folder and it goes, reference all of this. And you can make folders in chat GPT to where all of your chats about Project X stay together to keep it nice and organized. So yeah, you could, it's just a little bit more of a roundabout way. It's less plug and play than a skill. All right, so what I have open here is I have Codex. So this is kind of equivalent to the Claw desktop app for the Claude Claudeys in the chat.

1:12:19And oops, so basically what you would do is down here, I'm going to zoom, zoom, zoom, down here under thumbnails. This is just a pro, this is a folder that I have on my desktop. As you can see here, I've created the neuron general purpose, which is very similar to the one on chat GPT, but you can say add new project, start from scratch or use an existing folder and basically what you want to do is you want to just download all of the things that you had from like all of the knowledge that you've created in your mobile project um i keep saying mobile in your web-based project or your chat gpt-based project let me pull this over here to compare so this is very zoomed in i don't know how to get out okay here we go so right here i have my project folder here i have my project folder i'm going to click on this i have all of these chats okay i can maybe download that that would be kind of tedious so this is where it's a little inconvenient but i could also just come here and download all of these into a local folder and then that local folder i say use an existing folder and then i would like let's say this you know this is my my local folder the neuron general then i would just open that and that's boom now i have essentially the same thing the only thing you would have to transfer over is if you have specific instructions as part of the project folder and we cover all of this in our ai for total beginners so if you're like whoa whoa wait a minute you didn't tell me projects have instructions go watch that yeah um and uh and then you can copy all that and then i don't know where you would put instructions here.

1:14:06I think what you would probably want to do is you would probably just save it as a local... Like a desktop folder or... Yeah, Word document. Or you could even say it, like, hey, here are my project instructions. Can add these to a local file called instructions. That's not it. I didn't copy it. I wouldn't call it instructions either. I'd call it this project's instructions or something because you might have a bunch of them called instructions for you're over yeah exactly um i hope that answers that question um i thought it called you an octopus i've got to put my glasses on did this what it says on the main screen there i read outputs as octopus the only thing i'm wondering is if you could share it with your codex self no you can't um i think the super app that they're creating is going to solve this problem though i do too i do too i think there's a number of things that aren't aligned between codex and chat gpt and that it really benefits them too because every one of those is eating its own compute it's uh it requires separate people to manage and they're very similar but not the same and i think they'll line all that out as soon as June.

1:15:24Yeah, I think so. That's the rumor.

1:15:31So, plugins. As you can see here, there's a plugin for Chrome. There's a plugin for GitHub. These are things you might know. But then there's something called beta analytics. What is this? Okay, we're going to click into this. now you can see okay this plugin actually contains 20 different apps so these are the 20 different apps that you might need if you do data analytics on a daily basis right so it has these apps now these apps are technically mcp servers um or what you could refer to as connectors so these are apps that as you can see here some of these i already have installed like i have slack installed i have Notion installed.

1:16:15I don't have Google Drive installed here, but in theory I could. But then when I add this plugin to my codex, it automatically installs all those for me. Think of these plugins as exactly the thing in Claude that broke the stock market a few months ago. When Claude released the legal plugin and every legal stock crashed, Claude released another one a few weeks later and they all crashed. These plugins are the same kind of thing. You can come in here and do a lot of data science stuff, you might have needed a dashboard tool or a major BI enterprise tool for. I mean, and, you know, depending on what you're doing, how big your company is, you may still, but you also might not.

1:16:58So here's exactly what Corey was talking about. So on Claude, for all the Claude-ies in the chat. So you go to customize and then there's skills and connectors, but there's also plugins. So you can go to browse plugins and it'll show you all the plugins and it's the same thing so this was the one that broke the chat again i had the kind of moment where i was like well i don't get why this is a big deal like all they've you know you could be creating skills and and connectors before but basically this just puts it all in one place so as you can see here it comes with nine different skills already built in and then you can click on codex has a baller one called build ios apps yeah yes um but it shows you all the different skills inside here and this is how it works.

1:17:42It's like, oh, you can do legal risk assessment. You can do a meeting brief. You can review contracts. You can do a triage and NDA, a vendor check, things of that nature. And that's all just built in. So the way that I would think about a plugin is once you've got a lot of skills that you use related to a project and you have certain connectors that you've added related to that project, you can create a plugin that puts all of them together in one place and then you can share that with other people on your team. Like, so you could say, hey, this has been a system that's working really well for me.

1:18:18All you need to do is add this plugin with this little plus sign here, create a plugin. You can add it to the marketplace or you can share it with your team. This is not a business account. This is the free one. So we're not seeing that option here. And then you can make it accessible to everyone. So you can think of plugins as like the place where you organize your skills. So skills are like the place where you organize your prompts and your chats and your workflows, plugins are like the place where you organize your workspaces. Sort of like, now we're getting into like, it's like a project. It's an organization.

1:18:53There's a lot of gray. There's an increasing amount of gray as these tools grow. Yes, which is why I wanted to do this three, because it's all confusing and we need to make it simple. So you can think of plugins as a place to organize your connectors. They unlock so much. Yeah. They really do. Like, skills can unlock so many things. When I taught this, our sales team sat down as a group and came up with 100 ideas. We estimated that that meeting with 60 people in it saved well in excess of 1 ,000 man hours a year right then, like in the room. And not to mention what they've learned since they got back, how they've implemented them, sharing them with their teams.

1:19:35So there's so much potential there. And I know I'm talking about it from a business aspect, but they're just as valuable in your personal life. Maybe what you want is one that knows what you'll eat. Here are the restaurants I like. Here are the things I cook. Pick me something for dinner. There's a lot of things you could do. I've determined. I've said since the beginning of AI that one of these days I'm going to find an AI solution to end the nightly dinner debate. I don't think you've got it because it keeps coming up. I don't. To be fair, I'm confident I could now. It's more the laziness around creating a list of what I do and don't want to eat.

1:20:13Yeah, yeah, that's fair. Hey, going dark for just one second, Grant. Sure. I'm going to double click on something that Corey just said. Sorry to use a very corporate term, but I'm on a computer, and I am going to double click on something. So here, for example, this is a plug-in that I believe OpenAI created, and it's to build iOS apps. And so here's another example. So it has about nine skills built into it. So you don't have to go track down skills and some sketchy skill library on the internet. It's like, these are all built in here. And these are all presumably skills that, who created this?

1:20:50Yeah, you can go in and you can check the website and see who created it. It shows you, it's okay, this is connected to the Xcode build MCP. So that's related to, you know, not to get too technical, the Xcode is for, you know, developers with Apple developers. And yeah, it has all these skills. And you can now add this plugin without having to manually figure out how to do all of these workflows yourself. You can just add this and then you can create iOS apps with it directly inside Codex. So it's pretty cool. That is how powerful it is when you start bundling and adding all of these things together.

1:21:28And there's one other thing that I wanted to talk about. And I'm going to example it here, and that's connectors. So we sort of glanced over this. We go a little bit more in depth on this in our Agents for Total Beginners. But basically, connectors are the way that you can add outside applications into your AI workspace. So you can think of connectors as the way that you can get your AI inside of ChatGPT or Cloud to talk directly to, you know, other applications. So in this case, these are some examples I've shared before, but I've got it connected to Gmail. I've got it connected to Google Drive, to Linear, which helps me manage, you know, code projects, to Slack, and some of these fun things here.

1:22:12I'm the MVP. Oh, thank you. Thank you. Thank you. I appreciate it. I won't tell Corey, don't worry. And I do like the way that Claude breaks this out for all the Claudeys in the chat. So as you can see here, when you go to customize connectors and you see all of these tools listed out, all of these have permissions that you can set for any time that you want to use these particular skills. So it breaks it out into read-only tools and write and delete tools. and here you can go in and you can say okay you know maybe i'm gonna always allow maybe i'm gonna you know this needs approval or blocked typically with read only i'd be like sure only always allow for most of them and then maybe the write and delete ones yeah this needs approval because i want to make sure i know what you're doing before you go do that yeah right and i like the way that this uh they break it out in claude it's very uh granular and very user-friendly and kudos like It's like permissions in an API.

1:23:15Yeah. You know, when you're setting up an API, you know, you can turn on and off certain features, certain tools within the API. It's kind of the same thing made simple. Does Codex just list everything under plugins now, even if it's a connector? In Codex, I believe so. I believe it's plugins are connectors. Yeah, they're kind of the same thing. Yeah. They used to call them connectors in chat. I believe both tools now have a plugin creator. So basically you can just use that skill and put all of your skills together in one place, which is nice. One thing to know for sure, whichever one of these you use, if one of them has a cool idea, the other one will have it in four weeks.

1:23:59Yeah. They're pretty good at... Borrowing liberally from each other. Yeah, ripping each other off. straight up stealing for each other yeah yes um that's funny okay so that covers guilty parties yeah sorry for giving the parallax here i'm going to pull my screen down for a second so that covers all of the basics right so we've got um you know custom gpt's we've got projects we've got skills we've got agents which we touched on very briefly we're going to go a little more a little more in connectors. So that's six different things. Shout out in the chat if you understand when and where you would use all six of those things.

1:24:47And don't worry. And if there's anything you don't understand? Yeah. Ask away. Ask. We can bring them back up. We can talk you through something. We're glad to answer any questions. Whatever we can do to help you leave here today with more knowledge than you arrived with, we are down for. If not, there is much drama in the AI space this week we could always ramble about. Exactly. While I'm waiting for your comments to trickle in because there's a bit of a like 10 to 15 second delay, I'm going to give you a simple decision tree. Okay. So we introduced some of this, but we introduced some new concepts that we're going to add to it.

1:25:23So do you need a place to organize ongoing work? Use a project, right? Project folders are the place to do that. It helps you stay organized, organize your chats. They nicely get folded under the project folder so you keep it all in one place. Do you need a reusable assistant for a role? That could be a custom GPT or JET. Could also be a workspace agent, right, if you have access to those. Do you need a repeatable process? Use a skill. The reusable repeatable process, if you're going to do it more than twice, it should be a skill. If you need an AI to do actions across tools, so across different projects, across different tools, you're going to use an agent for that.

1:26:06And if you're using Codex or you're using Cowork, you can think of Codex and Cowork as the agent. So you don't have to build separate agents for that. Of course you can, but you don't have to. You can use Codex as your base agent. You can use Cowork as your base agent and you can work off of that. Now, we haven't talked about subagents yet, so we'll skip that one. But if you need AI connected to apps, files, or external systems, you can use connectors or MCBs. So if you need your AI to connect to Slack, if you need your AI to connect to Google Drive, if you need it to be connected to some other tool that you use, that's where you would use a connector.

1:26:49Do you need to package the setup to be reused? that's when you would use a plugin. Or you could use somebody else's plugin where they've already packaged it all for you. So you get access to all of these skills that somebody else took a lot of time to build, which is really a good unlock and a good way to get up and rolling with stuff very quickly. And do you need a system to keep checking, fixing, and trying again? Then that's where you would bring in what's called a loop. And we can get into that in a minute, but I want to make sure everybody understands everything that we've talked about so far before we get any more complicated.

1:27:27I'm dropping the little notes. Oh, I love it. Thank you, Corey. In here for folks, because it feels like a thing that would be good to put on a post-it at your desk. Yeah, we're dropping everything I just said directly in the chat so you can access it. Please know the formatting has got awful. You'll need to fix that. Just drop it into your chatbot and ask it to make a neat bullet point list for you. or turn it into a graphic. Like you can use the image generator to do that. You could do that too, couldn't you? Matter of fact, maybe we'll make that into, I'll use my skill to make that into another graphic you can share out as well in the newsletter.

1:28:02There you go. Please do. I love that idea. Does it go good in AI skill of the day, maybe or something? Yeah. Yeah, we can use it. All right. So I'm going to talk about loops for a second, but do, do shout out in the chat. Please. Nice. Please shout out in the chat if you want follow-ups on anything. So a loop is a repeated process, right? So this is basically where you're not taking yourself out of the loop per se, but you're setting a loop into motion where an AI can basically do all of the things that we've been talking about on its own without you having to come in and prompt it. So the AI checks its state.

1:28:47It takes actions based on the state. It verifies whether or not those actions achieved the goal that you're trying to set out. It improves and tries again until it hits a stopping point. So you might have heard of this really cool feature that was just added called backslash goals. Backslash goals is basically a way to set up a repeatable loop for your AI to go off and try and accomplish a goal for you. And the way that that works is you set up a goal. You say, hey, I want to build an iOS app for kids. Oh, yeah, I love Corey's graphic. Let's do this. So what does it say? Oh, wait, wait, wait. I got things.

1:29:29What does it say? Yeah, so I did this video trying to explain the difference in a prompt and a loop. You know, essentially a prompt is an ask once, answer once thing. Where a loop, you're like, I want you to do a thing. The AI goes, does its work. It comes up with an answer. It's going to validate it, and then it's going to come back around and get recommendations from that validation and keep going. And it will do it over and over until you've told it to stop. If you read the neuron and remember us talking about the Ralph loop a number of weeks ago, loops have problems, okay? The other thing to know is that, you know, it doesn't hurt to tell it, do this twice or something.

1:30:11Developers use these a lot in areas where they're cleaning up code, where they're doing reviews and things like that, trying to make sure things are efficient. But there's definitely... And there's a really great blog related to this that I'm going to include in the chat that I don't want anyone to get too distracted by. But basically, it's from Addy Asmani, who I believe is a Google engineer. I could be wrong about that. Talking about how you use loops in an agentic setting. and one of the most important things is making sure that you have what are called sub-agents and so we talked about them very briefly earlier but they're like a specialist worker inside of a larger task and this is where you might use an agent in addition to cloud code or to um sorry to code work or to codex basically it's a specialist worker that goes off and you know an example would be one agent maps the files one writes the fix one reviews the files for security holes and then one checks it against the docs and it's like, did it actually work?

1:31:11Now, a lot of these tools, these tools are getting so good that it does this automatically for you a lot of the time. But sometimes you have to still prompt it and say, hey, create sub agents for this. Sometimes I have a series of product team sub agents for when I do for when I build. So I've got one that's like a UI designer. I've got one that is like a product manager. I've got one that specializes in the user experience and one that specializes in back-end design. And between the four of those, like, you know, anytime I build anything, I can be like, run the product management suite. And it goes through each of those agents and comes back with a report that it has read, condensed, combined, and fed over to me.

1:31:56And then I tell it, go do that. And it goes and it just does that, which is such a sick thing that we really underappreciate. That's amazing. What Addy recommends is basically making sure that you always have at least two, which one is the one that writes it and one that's a verifier that you, you know, he's talking about it for engineers, right? in an engineering context. It'd be like, you know, make sure that you teach the verifier how to be very critical of your code, teach it what you wanted to, you know, what good looks like, you know, any sort of like specifics about your project that it needs to know.

1:32:31And the reason that you need two of them is because the one that writes its own code is going to think that it's stink don't stink, if you know what I mean. Like it's going to think it's great because it wrote it. So you need a second one to basically come in and verify. And I saw someone in the chat mentioned earlier, they use Claude to verify ChatGBT's work, and then they use Gemini to verify Claude's work. That's a great way to do it as well. Subagents is just doing that all in one chat, in one process. And they all have built-in tools, or at least Codex and Cowork have built-in tools to do this.

1:33:04But you definitely want at least a verifier and at least a writer to check each other's work. and um i think on codex you have to tell it like hey use subagents for this and claude sometimes does it automatically and then there's more complicated versions of this um but but yeah um you can you can apply it pretty simply we're really living in a time with an abundance of amazing tools available to us like just as recently as like december vibe coding a little tool like you had to go back and tell it fix this fix this you don't get it fix this and i mean like that thing i built last night i i spent 10 minutes i was one prompt if i was any any things i gave guidance i gave it after that was you know what i'd like to add this i'd like to add this feature additions it was not yep this doesn't work this isn't doing this it just worked first time the worst part about this era of building like software applications is feature itis like you just start thinking, well, I could make anything.

1:34:09So why don't I add this? Why don't I add this? The hard part is being reductive and taking stuff out. I saw a message in the chat I want to comment on. And on the same page, some of it requires practice, skills, and agents to understand, but things like connectors are used in other apps. So the penny drops on those. As of the first, I've heard that phrase. I don't fully follow. Yeah. If you could, if you could, as a dear friend of mine says, say that again with other words. Yeah. But I like it in principle. I like it. Yeah. So the way I would, the way I would think about skills and agents is I think you should try, you should try to go out and create a skill for something that is a process you do every day.

1:34:53And that can reasonably be accomplished end to end by an AI. So if you think, you know what, if I just give it, you know, these constraints, this goal, these instructions, and, you know, it has to be formatted in this way. Maybe it needs, you know, this code or that code in order to make sure that, you know, you're checking the data properly or something, if it's a data task. And you can get that to a good point where that process can be turned into a skill. if you can tell it what you want and reasonably expect to get it back it should be a skill and you should try it like today and try and go make one because and and and like like the instant each of us figured out skills like Grant got them before me and and the instant there was a click moment it was like oh my god and I'm running through my chat archive and I'm like make it a skill yeah make it a skill yeah you get a skill and you get a skill and But it was one of those pivotal unlock moments with AI that was just a feature tacked onto a model.

1:36:04It's not AI. It's using AI. But sometimes that is what really matters. I want to bring up this other question here because it's a really good one. Do you ever encounter runaway loops where the exit condition is never fulfilled? When I mentioned a few minutes ago that there are drawbacks to loops, that's a thing. And it'll depend on your problem. Like if you have a thing that you're like, this is going to have to run for a long time, running long is not necessarily good. Because there's a cost associated with it. Yeah. Yeah. In my limited experience with them, I'm still learning my way through them.

1:36:51I often find myself setting a qualifier, like, do this three times, where the exit is not subjective or conditional. It's very quantitative, just so it knows, okay, I'm going to do this, then I'm going to exit. And you know you're going to land better than you did than you would have otherwise. But you're not going to let it run so long that it beats a dead horse or whatever the saying is. I will say in my experience, I've not had this happen. But I am very, I'm still pretty manual with the type of agents and loops that I do. I'm pretty much restricting my loop usage, my loopage to using goal.

1:37:36And, you know, again, that's backslash goal. I'll type it in the chat. and goal's a loop and goal yeah goal is a loop but it's a loop towards a purpose and at least in my experience with Claude because I'm doing this primarily with Claude when I'm doing this I find that it will often under complete the task before it takes too long working on something if that makes sense like yeah i find that i'll i'll come back to it and it will be 90 of what i specify good enough or yeah or yeah it'll it'll meet the specification criteria because the thing about goal so we kind of touched on this but let's just make it very explicit when you set a goal it's like giving a prompt where you're asking it to complete something but then related to that you have to clarify how it's going to know that it's done.

1:38:30And so you have to specify what the exit criteria is. So you have to say done looks like this. And a lot of times when I'm working with it, which Boris Churny, who's the creator of Cloud Code does not do this, but I will create a plan first and I will say done looks like this plan fully implemented end to end. So this is what done looks like. If you've done this plan, it's done. Right. And that's the way that I do it. but Boris apparently doesn't do this at all and just uses goal and that's it and in auto mode yeah which is that's where that's for engineering so that's probably like a little more complicated than most people here need but I find that it often will finish before it's truly done before it will spend way too long running and going off forever but that's also keep in mind that is me using it on the subscription tier that's not me using it on the api they might be shady they might they might cut off the users more in the subscription tier than they do the api tier so yeah perhaps if you're using it for the api and what that means is you're paying per token and you're you know if you're applying this in a use case where you're actually like running the agent as a product or something like that um you're gonna want to be a lot more careful about that and do something like cory said like where you set like okay if you've done this more than five times stop well five times stop you've yeah you've probably we've reached the point of diminished returns uh you know uh yeah if we're at 98 don't go fighting for 100 yeah for 99.6 don't keep fighting you're there you want you can also add yourself as a in the loops and which was probably recommended to do um so you could say like after you've done this three times you know give it give it to me on slack to rodeo or something like that um and that way you're setting the exit criteria so that it's set up yeah i want to pull this up because this is such a killer killer point the model can already pick which skill to fire once a skill exists what it can't do is notice i keep solving the same problem by hand and tell me to make it a skill or an agent for the love of god somebody build this in anthropic open ai gemini anybody somebody build this so the others can steal it too and and put this in your tool because that is a thing i would love to be able to be like oh you know what you're right one click i know it's been tell it where it's just like hey you've done this like 17 times would you like a skill and you hit sure make a skill and it does it that is the thing we we really need that it'd be a I mean, it's not like it's a big productivity saver, but it is just another thing to remember.

1:41:17Like, you know, I ought to make this a skill sometime. And then you just move on, you forget. And then tomorrow comes and you do it again and you're like, dang it. Okay, so my first thought when I heard this was, oh, well, the reason they haven't done this yet is because their memory systems suck. Or at least by suck, I mean they can't actually keep track of everything, right? Like they're not perfect. Like the count of things. Yeah, yeah. They're not perfect, right? But they can search your previous chats. Now, if you're using projects, you have to be careful with the way you set up settings because you had a realization that seemed very exciting.

1:41:53I'm trying something. Okay, you do that. Let me finish my thoughts before I get distracted. If you're using projects, you have to be careful with how you set up your settings because sometimes the search will only refer to chats in that project. but let's say you don't use projects and you're going YOLO mode. You could actually create your own skill that does this. You could say, hey, every time I ask you something, check to see if I've asked you for this before in my previous chats. And if I have, suggest that we turn it into a skill. And you make that the skill that automatically fires. And then it will do a search through your previous chats and it will look and see, oh, okay, like you've asked me for this like three times.

1:42:38Maybe you say like instead of more than once you say like three times or five times whatever your Criteria is and then it will automatically prompt it And then you set that skill up so that it runs no matter what you ask You basically just tell it like at the beginning of every chat run this skill And you can set that as a project instruction At the beginning of your project and I'm sure I'm going way too fast So we'll write this all in the blog that goes with this Or you could just try to set it as like a global memory feature depending on which tool you use Someone asked, how do I delete chats and projects to make better manage my files?

1:43:13We could show that. Yeah, I actually had just sent an answer. But that's a good call. My answer was, unfortunately, one at a time and manually. Yeah. You guess you don't really want to give it ability to auto archive. I will say if you are going to switch to codecs, one of the things you could do is you can actually set up. automations that run and automations if they run and they fire and nothing nothing happens because the criteria wasn't met they'll actually auto archive so that could be one way to keep your chat messaging a little bit more under control as you could set like okay we're gonna you know run this you know thing that i ask you for every day or multiple times a day i'm just gonna run it on the hour and if nothing meets the criteria with which you know like let's say like every time i add a new row to this spreadsheet, you know, process it.

1:44:11Or every time I add a new file to this folder, process it. And, you know, if you haven't added a new folder, then it'll just auto archive. So that's one way you can kind of keep control of your chats. But yeah, it's manual, unfortunately.

1:44:26I saw another comment here. Is that meta layer coming or is, oh, that was, we already addressed this. Plugins and connectors are used in other apps. Google Sheets has connectors or one makes the connection but agents and skills are new to me i understand the concept and the analogies interesting do you have any response to that core

1:44:53i missed it i was deep i i was deep and uh absolutely missed the comment where is it i don't see it uh so it's a Andre Day. Is this plug-ins and connectors that are used in other apps? Yeah, Google Sheets has connectors. Yeah. Agents and skills are new to me. Okay. Cool. An analogy is out, but if you have to use them, give it a try. It's one of those things that agents feel intimidating when you first start. If you've ever used a tool like Make or any kind of workflow automation tool, agents will be a breeze. If you haven't, expect a minor learning curve. Don't dive into the deep end. Use a tool that's made to make it easy.

1:45:42There are tools now that do natural language agent construction, and that's absolutely the way to go. Like, you know, if you're experienced, maybe you're using an 8N or something. And I once did back in the days of yore, you know. Back in November 2025, when the world was a different place. But now I'm primarily using workspace agents or a variety of things like that where it's just easy to say, like, here's what I want. Do it. We got another question. Sean Deadman said, great show, Corey Grant. Can you automate the savings of Project Work folders to external storage resources or did you already cover this?

1:46:30Boy, that would be great. I have been hanging on to this idea. And you know what? We'll just say it. We saw a thing at some point in history where there was kind of an implication that Chachibiti might at some point add a drive-type solution, cloud thing. To date, that has not materialized. It's still a thing that I think is coming. It's going to put them in direct competition with all of their cloud providers, though. So it's definitely complex You mean this?

1:47:05What are you doing? Are you doing... Oh, okay This just happened today Oh Help organizations Deploy agents for it Securely Execution taken longer running work I don't know No I was I was talking about A report You and I No, I know what you're talking about, but I mean, like, this is interesting, and I wonder if it's related to what you were just talking about. Good question. Good question. Spence Kodish for Secure, customer-controlled cloud infrastructure for long-running agents across software and knowledge work. Just saying. Could be. It definitely feels like a piece of a puzzle, I would say.

1:47:54There are a number of things that since then have looked like, oh, look, that looks like part of that. Yeah. You know, like the fact that Codex has its own Sheets plug-in and its own Dux plug-in and its own Slides plug-in. You know, I mean, it definitely has the Microsoft and Google versions, but it also has ChatGPT's own of each of those. So I think related to the question. That's what it's going to take to have organized project work is what I was, what I should have. Yeah. Yeah. Yeah. That's fair. Well, I was going to say related, related to the question, right. You can do this. So automating the saving of project work and folders to external storage resources.

1:48:41So one way you could do this is you could connect your Google drive. So I'll share my screen real fast. Okay. Okay, so in this case, I have the drive connected. That's not how you access them on Claude. Okay, connectors. I guess I can just ask for it. Like, can you add this to my Google Drive? And then maybe I'll just take the text of that article I just read, paste it in. And then essentially what it's going to do is it's going to find the tools. It's going to ask me for permission because I haven't set always allow yet. I'm going to be like, yeah, why not? And then in this case, it's going to then create the file on my drive.

1:49:30So in this case, I guess I'm technically automating the creation of work and folders to external resources. And, you know, in theory, I can connect both, you know, both tools to my drive and could then access all of my resources there. I can save all my skills to my drive. Then they're all in one place. So all I have to do is just say, hey, grab this skill from my drive. You know, in this case, it's asking me to connect. So I'm going to do something with that. You know, something I want to see is I want to see OpenAI connect the library in ChatGPT that has every image and every document you've ever uploaded to your codex as well.

1:50:10I want to be able to access one library from both places and have anything I input be saved accessible from each. Adding codex capabilities to mobile was wonderful. Like I said, I test drove that this morning for the first time, and I felt like, I don't know, like, you know, Mr. 3000. Not Mr. 3000. It was a baseball movie. The other way I would do it. I felt like the Terminator. How's that? I felt like I was doing magic in the coffee shop line this morning. And to know that, okay, hit the thing, it works. It was essentially what I would do if it was OpenClaw I was working with. The other way I would do it is, so you might not be able to, let's say, automate the process of creating project folders inside of Codex.

1:51:01But once you have a project folder inside Codex, what you could do is you can then ask it to create project folders on your desktop because that is access to your local file system on your desktop so for instance here you can say like go create a bunch of folders for me to manage all my work related tasks in a hierarchy and then it might ask me some follow-up questions about that but basically that's one way i could do it and then And I'm at least managing the folders on my computer as another external resource. So you can do it with Beehive. Sorry, not Beehive. You can do it with Google drive and you can also do it locally on your computer.

1:51:44Yeah. Yeah. Very cool. Very cool. Anything else that we didn't cover before we wrap up?

1:51:56I think that's it. I really do. I think that's it. I think we've hit all the questions. I really appreciate that looks like 50 of you have managed to hang out for two whole hours that's impressive we always look to see how many are left at the end, that's how we gauge whether it was a successful live or not not how many show up at the peak but how many hung in there for two hours and I'm always really impressed with how many it is I can show an example of goal real fast before you drop that's cool if you could let my external computer join for a moment I can, I can Do you want me to kick this one first?

1:52:36Yeah, you can take this down, that's fine I just wanted to show the demo of how you do that I'm going to be honest, I'm not I can take it down Yeah, you take it down, because I'm going to kill it or you And I'm not sure which Yeah Alright, so You should not be seeing anything yet But that will change in a moment

1:53:00Let's see.

1:53:08Okay, I'm gonna have... Okay, I think I have everything set up, so I'm gonna go ahead and share my full screen here.

1:53:23Okay.

1:53:28so in this case let me know if you can see this because i can't see my screen here i can okay so i've been working with the new fable five did y 'all hear about this um maybe you read about it in the neuron it's like the new uh claude model that was very controversial when it came out and i recorded an ugly short that claude fans won't like yesterday i say ugly i was nice but i was but basically i made it so you can customize a little squirrel that you make um you know you can push your tail perky ears um i guess i'll use a passphrase well actually let me see if i can do it without a passphrase okay no i have to do one one two three um so i made basically like runescape for squirrels let me see if i can zoom oh i can't zoom it in at all but yeah basically like so this is a little squirrel um game and then this is like the town and you pick up acorns and um in this case all right i'm walking around and i'm checking out things and subtle nod to michael phelps and as you can see here like okay now i'm going up the stairs to the tree um and what is this oh i could interact okay oh it's an anvil okay i can i haven't learned this profession yet so you know etc etc so you get the idea so now what i can do is i can go to claude code and i uh give me something that we want to add to this game at literally anything uh the ability to spin the screen so you can see around that tree give me the ability to spin the screen and i'm going to add and zoom in or out yes so i can what was the what was the phrase that you used see around that tree see around the tree and different assets right um and i'm gonna say um complete looks like um what i want to say an end to end 360 degree view that I can

1:55:42use and switch around at any time. Now, it might have a problem doing this because

1:55:50isomorphic games, basically, they render in a certain way, so like it's from the point of view that you're looking at, but in this case, I'll say I can switch around at any time and zoom in and out that is usable in the game, right? And in this case, it's actually already running something. So I'm just going to cancel that, and I'm going to say this is the goal. So in this case, as you can see, I use backslash goal, and it's going to set up the goal. It's funny that it requires a goal. It doesn't need to. Because to me, it feels like I've prompted you to do this. This is your goal. Yeah. If I hit enter, this is your goal.

1:56:29Goal is more about the looping process, right? Okay. So in this case, it's about making sure that it has a heartbeat and that heartbeat continues to continuously work on this until it's actually done. And you can get really specific with what complete looks like. Like I said, I usually make a full plan and say work on this plan. In this case, this is just a feature that we're going to add to this, right? And so it's going to go off and it's going to start working on that. and we can see this will update in real time when it's actually implementing it. Now, it didn't immediately come back to me and say, hey, that's not going to work.

1:57:05So in theory, it's going to try and do it. And I think I'm working on, let me just make sure I'm on Fable. Yeah, I'm on Fable 5, so it should do it, right? Extra. Yes, extra high. That's the - Rate limits be damned. Well, I'm on the 200 plan on this. So you'll get some decent use then, hopefully. Yeah, yeah. I see it's not going to remain on plans. No. What do you mean by that? Sorry. It's being removed from subscription plans as of it was the 22nd. Now it's the 30th. Yes, this is actually something I want to emphasize for people. Use this model now because it is never going to be this cheap to use this again.

1:57:49If you have something that you want to create, now is the time to do it. because otherwise they're going to jack up the prices. Did you see OpenAI's response this morning, what they're doing? No, tell me about it. There was a report in the Wall Street Journal that they're looking at drastic rate cuts in response to Claude's high rates. You know what? I think I did here. I'm like, as a consumer, I love this plan. As a consumer who's not an accredited investor and can't buy in, I love this plan. By the way, if anyone wants to actually play this, I can make it accessible and it's supposed to be able to host at least a hundred people at a time so I can post it on railway and and we can play it maybe the next time we get together there we go but I don't know there was a dungeon so the first version of this was a dungeon and now this one is just like the village and I don't know how to get back to the dungeon I gotta ask it who's this this is an npc okay I can buy oh red berries hazelnut paste interesting Ooh, hazelnut paste.

1:58:52Yeah, let's up with that. And then I also want to make this, so this is like a platform, but I wanted this to be a bunch of wood bridges, so I'm going to have to give it feedback on that and say, like, hey, this kind of sucks.

1:59:09In the grand scheme of things, it doesn't. I saw someone, like, made a really nice Minecraft clone the other night. Oh, yeah. I feel like you could pretty much clone anything with the same style. And to be fair, Minecraft is like the most clonable game of clonable games. I would think, um, in a, in a very Tetris kind of manner. Okay. So I just accepted this quest. I don't, no idea how to actually find out if I'm on it or not. I like it. Oh, you know what? I'm seeing a little purple. Maybe you should go to that purple thing on the mini map. You see that? Maybe that's your, uh, yeah that's something you could add you could have it add an arrow around the outside of the mini-map pointing to where you're supposed to go or a dot on the mini-map for your goal in yeah oh here we go obvious oh okay so now we're in the cellar right so now it's a party yeah so i can hit space bar to attack so i'm gonna go let's see here

2:00:14these spiders hurt yeah zoom in would be nice that's definitely a thing that's missing yeah i can't zoom it can't zoom into this you're like i'm hitting buttons and like things are flying but okay it's continuing the camera where it says done 360 and pushed do you need to reload Yeah, probably. Let's try it. Okay, Michael Phelps. Oh, what was that feedback we just had for it? Let's give it that. Mark goals on the minimap?

2:00:49In an obvious color? Like, you know, red or flashing or something. I think WoW uses red with a ring around it.

2:01:01fishing, flashing It's so weird glancing at the other monitor and seeing you minutes ago. I'm going to say goal. It would be nice. Also, the different layers of the, what's it called? Tree bridges should be walkable underneath. underneath like a true bridge maybe add some more trees have you used anti-gravity and what is your experience with it grant have you used anti-gravity a little bit i haven't messed with it much to be honest um i played with it touched it but i i but i will i i should because it's a thing that's been on my list okay all right so that's set so now let's go see if the camera work worked um so one two three okay oh a squirrel oh oh okay i need another option which is hey how do you log back into your squirrel once you log back return to game yeah we should add a return to game feature yeah all right so i'll add that to this so anyway let me do wiggle phelps 2 Okay.

2:02:30So now, how do we scroll to Zoom? Okay. Yeah. Oh! Kicked me out. Gotta go back to Michael Phelps 3.

2:02:45Okay. Your middle name is Bowman? I didn't know that. Well, no one said that technically. Okay. Okay. so I can zoom in yay alright so now you can see the squirrel a little bit better nice oh yeah that's helpful he's a fat one yeah a little chunky guy I like it alright so I'm getting my acorns getting my acorns alright are acorns experience points? yeah yeah I completely made that up I would think acorns would be like you're healing yourself whoa they're coming at me oh yeah you got problems bro what is this oh i would stop running and like aggroing them all into the room with you yeah well i'm trying to like gather some stuff but i guess i need the profession to do it what what is your weapon here uh i'm just just my little claws at this point let's see i'm going to i need to level up so i can get some experience Oh, now you ran into more of them.

2:03:56Yeah, I'm just going to have to outrun them. He's just going to Leroy his way right out. Yep, I just got to keep going. This is what squirrels do in real life, you know? Uh-oh. 10? Ah! Maybe I should add, like, a limit to how far they can chase me. Oh, no. Yeah, they have been still coming. They also see you, like, from so far away, which is really frustrating. but I am getting my health back as you can see as I run away. Maybe you need light. Okay, let's see if I can take this guy out. Actually, you had a pretty fair clip. Ooh, I got a dagger. Okay, great. Let me try and take this dude out. I'm going to stab something.

2:04:37Stick, stick, stick. Whoa, you're dying pretty good too. Oh, I got squashed. The Dre calls you home. I don't know what that is. Grant, that was a ton of fun, man. To everyone who joined, I really appreciate you taking the time today. Please Take just a moment to like and subscribe to the channel. It helps us out a lot. Also, pop by the Neuron.ai. Sign up for our newsletter where you can keep up with all of the nerdiness that you shared out every morning for your viewing and reading pleasure. Also, huge thanks to Beyond Trust for sponsoring today's live with us. Really appreciate it. Make sure you go take their identity security awareness test and see what you think.

2:05:19We're going to drop that link right back in the chat here. It's pinned. It's pinned. Oh, it's pinned. Okay. There's a link pinned below the thing right down there. Click it. Go to Beyond Trust. Check them out. They've got a cool thing going that's really worth looking at. And as always, we will turn this into a helpful guide that goes along with this live stream so that you can follow along if you missed it. and it will be archived in case some of you are not monsters who stayed for the full two hours and would like to go back and catch the beginning where we actually taught the things. But if you want to play squirrel, whatever this is squirrel game, let me know and I'll make it accessible for people.

2:06:01There you go. And make sure you go watch the Mustafa Suleiman interview we did that just released yesterday on the channel. It's really good. It's only like 15, 20 minutes. and he dished a lot and talked about some things that are really interesting for AI development that they're doing that I haven't heard of elsewhere too much. So that's it. And we don't have anything else. So on that note, farewell for now, humans. Have a great weekend. Bye-bye.