Transcript · 26,518 chars
I'm going to show you how to make Chat GPT into an agent that will do your work for you. It can read and reply to your email, it can book meetings for you, it can schedule events, it can create gorgeous ads for your business, it can act like an employee that actually does work. Now, I always thought of Chat GPT as the dumb chat app, but ever since they merged it with Codex, it took on superpowers. To master it, I watched hundreds of hours of YouTube videos and understood what works. I'll show you the clips that I think are the most useful for you, and I'll have links to everything below. Let's get into it. Presented by Zapier, the AI automation company. Here's the first video. >> I always found it a pain to open Google Calendar and manually schedule an event by clicking a bunch of buttons and entering a bunch of text. It's much easier to just ask Codex to book calendar meetings for you. So, for example, I asked it to book a Google Meet meeting between me and Char for Friday at 9:00 a.m. PST called Peter Char for 15 minutes. And Codex basically went off and it found Char's email, right? And it even found that I had a conflicting meeting where I had to drop off my kid at her camp during the same time. And eventually booked the calendar event here and even added a Google Meet link. If you're booking a meeting to, for example, meet someone for coffee or lunch, you can just mention the name of the restaurant and Codex will find it for you. Or if you're having like just kind of like a WhatsApp conversation with someone and you agree to a meeting time and place, usually what I do is I just screenshot a conversation and copy it directly to Codex and it's able to book the meeting. >> I love that he did that, that he just took a screenshot. Now, the thing that you need though is connection. So, let's take a moment and just talk about how to connect and get plugins in. >> It's plugins and skills. These are both under plugins on the side panel. And starting with plugins, this is how you connect work to external tools. There's tons of pre-built connectors, all the common ones like Gmail, Google Calendar, Drive, Canva, Slack, Asana, and much, much more than that. Then there's some more unique ones like computer use and Chrome use where it can actually navigate autonomously with its own cursor and use the browser to complete tasks. Skills are preset instructions and files that let work execute a task in a repeatable way every time, like recipe cards for workflows. >> As far as I can see, there's no problem with adding as many plugins as possible. So, the thing to do is when you sit down, if you really want your agent to just make as many connections as possible. Uh let's go back now that we've made the connections and we've scheduled a meeting. Check out managing managing email. >> writing plays. One thing that I frequently miss with my emails is I read the email and I don't follow up right away and I just forget to follow up, right? So, what I did is I started by asking Codex, "Check my email from the past 30 days and list the top five open action items to follow up on and emails to unsubscribe from." And also link the actual emails. Do not share any confidential information because we're giving this tutorial. So, here's the list that Codex came up with. It looks like there is a new device logging. There is a Google payment issue. There is an automation that I built that broke. There is some thread about granola and a meeting with Brex, right? And here is a bunch of unsubscribe candidates. And if I click one of these things, it'll take me to the actual email. So, what's the next step here? Now that we have this list, what we can do is basically just tell Codex, "Hey, why don't you unsubscribe from all five emails?" And because Codex can use the browser, it can actually open each of these individual emails in Chrome and click the unsubscribe button and just get it done. So, it looks like it worked for 2 minutes and it unsubscribed to all five emails that I wanted to unsubscribe from. >> Yeah, and I like that it can take actions. The unsubscribe is a very simple action, but you can also have it draft messages and take other action like this. >> Use Codex all the time to find and follow up on previous emails. So, for example, I sent an email to book a hotel in Japan and they haven't replied to me yet. So, I basically asked Codex to find the email for the Ufuin Hotel in Japan to send a follow message, but let me review it first. So, it found the email and it said that I sent the last follow message a week ago and it went off and actually drafted a reply. So, I'm following up on my reservation. We want to stay during Christmas time, prefer this room, and so on and so forth. And because and because we have our Gmail extension hooked up to Codex, I can just tell it to send the email after I edit and review it. So, I just said send and it got sent successfully to the existing thread. Now, to take this to the next level, what you can do is you can train Codex to reply to emails using your voice and tone. If you look at this email, it's a little bit too formal. So, why don't we actually get Codex to learn my voice and tone for emails? So, basically I told it to review my sent messages and let's create a {slash} email skill to draft emails in my voice and style. And then it worked for 4 minutes. Let's take a look at what I actually did. It went through a bunch of emails and reviewed everything. It actually even recognized that my voice is different whether I know the person versus someone new, right? And then there's something here. And it basically created a skill here. Let's take a look. So, created a skill called email and here's the workflow and here is the voice that it should follow. Now, let's actually test the skill with the Japan email draft that we had. And 30 seconds later, you can see here that it drafted the email that's a little bit shorter and less formal, which is more casual like my >> All right, there are a couple of things that I say about that. I still do not want to check my email within a chat thread. I think it's just annoying. What I would want it to do is in the way that I handle my email is I have drafts created and I just read them in my email app. That way I can go through my email app as normal. I think email apps are well designed, much better than a chat for going through messages and I can decide to send it, edit, or or just delete it, or write my own. The other thing is that I think is just as important I don't like having my skills on the Codex app. I like to have them in GitHub in the cloud somewhere. That way all my other devices and all my other tools and all my other everything can use it. I don't want to trapped in Codex. I don't want to trapped on my desktop. If I have it in GitHub, then Cloud Code and whatever else comes out tomorrow, Hermes Agent today, gets to use it and I I like that sharing. Um All right. What you're going to end up having is a lot of these conversations and I I like how Peter, and we'll close it out with this, organizes his projects and threads and then we'll move on to the next person cuz I think it's also important for you to see how uh Codex can have an email inbox that it can check on its own, but let's go with to this first. >> Let's cover projects and threads. So, I have two pinned threads here and if I expand this, here are all my projects, right? And as you can see, I have projects for my personal operating system and various products and apps that I'm working on down here. Now, within each project, I like to organize threads based on individual tasks and workflows. So, for example, in my personal OS project, I have threads for making podcast episodes for various guests, for helping to edit my newsletter post, drafting and posting social posts across all social networks, and so on and so forth, right? And one thing that Codex does really well is it's really good at compacting previous messages in the thread while still remembering what they were about. So, basically, that allows me to essentially just have extremely long-running threads. You can see this is a very long thread with a lot of back-and-forth conversation. And I just keep going. I'm not worried about Codex getting confused or not knowing what's going on. Another benefit of of having one thread per workflow is that it is so much more easier to manage than having to manage a whole bunch of threads for individual tasks, right? Now, to organize your threads better, you can also drag and drop individual threads inside a project. And of course, you can pin threads to the top that you're working on. One more thing is if you hold down the pound sign on your Mac, you can quickly navigate between different threads. So, I can go to pound three, pound four, pound two, and so on to nav >> I love shortcuts. He said the pound, but I think he means the command key, which has a pound-like design on it. And you have to hold it for a few seconds for it to work, which is why I never knew it was there. I would just tap all the control, option, and command keys, never see any shortcuts, and realize they weren't there. I'm a huge shortcuts person. I love that I could do that. Okay. Let's [clears throat] talk about giving Codex an email that it can actually manage. I caught this little clip in a Greg Eisenberg interview with Dan Shipper. Check this out. >> He gives my Codex an email address that rides on top of my email. So, you know like you know the plus format in emails or it's like, you know, you can do Greg plus you know, codex@gmail.com. >> Yeah. >> So, it basically uses the plus format to register itself in email inside of your email. So, I have Dan plus Codex and there's like a random string so you can't just email it. Um and then I can give that, and then what what it does is I have this router thread that checks that email inbox regularly, every like 5 or 15 minutes or so. And then I can give that email out. For example, we have an internal agent in our Slack. I can give the agent that email. So, when someone asked me to do something because I'm generally not as good at like ops type type thing type things, they know to ask the Slack agent and the Slack agent knows to email my Codex instead of emailing me. >> This is brilliant. If this screen share doesn't make any sense with what he's saying, it's because >> [clears throat] >> it just happened to have come up in the middle of the conversation weeks and weeks ago. I didn't understand it at the time and now I get it. >> [clears throat] >> And the impact is like I I'm getting a cold, I think. The impact is you can have Codex check your inbox and know to only read and action the items that have that plus and whatever secret code you give it. And that is really uh effective. Um you can also give it its own inbox. The problem I found is that Codex and other agents will have a plugin for email, but it's only for one email account. And so, what I've done is I've used my sponsor, which is Zapier, with their MCP to connect a second email account. And so, I can say, "Check my email this way or check my email through the Zapier MCP." And that way it could check two different messages and it can do this for lots of other tools. So, maybe I not maybe I definitely have a Notion account that's my personal Notion, which I can connect maybe natively to these tools, and a Notion account that's my work personal work and now I have the ability to connect multiple accounts using Zapier MCP. I have that um and you can try for free by going to zapier.com/mcp. Okay, if you give it an email account, you need it to schedule >> a runtime. You heard Dan say that he does it what every 5 15 minutes? I forget what he said. Here is how to do that from the launch video that ChatGPT created. >> meetings. The next thing I wanted is for this to happen without me asking every time. >> [snorts] >> Let's try this. Can you be my chief of staff and give me a briefing every morning of what I have to prepare for the day and while the World Cup is going on, can you actually give me the scores of the German team and any other matches that I care about? And that's it. The automation is set and I don't have to remember to ask again tomorrow. And by the way, this also works on the go. The same chat is already on my phone. I can start on the web, leave my desk, and keep going on my mobile. >> And we'll talk about how to connect it to the mobile, but the scheduling is really effective. I used to think I had to go into the scheduling section and write a whole scheduling thing. I like now that what I could do is just what he demoed. When I'm done with a project, if I want it to be repeated, I just tell it repeat this at whatever time I give it, and then I can go into the scheduling section of the app and see what I've scheduled in the past. Okay, uh here, this is how to connect it to mobile. I think this is really important. >> The way to set that up is you want to go into settings down here under connections. You want to say control this Mac, allow connections, yes. So, you're going to click this add button. It's going to give you this QR code. You >> Which by the way, he hid the QR code because the cool thing about this is you can actually connect more multiple devices, not just your phone to your desktop, but a desktop to another desktop. So, I have one laptop actually that I take with me anywhere, and another one that I keep on my desk. I shut the one that I take with me, I still want to control the one that's remote, and then shut my laptop and have the remote work. Anyway, I can do that. So, you can control Codex desktop from multiple devices, and they don't have to be phones, and that's why and they don't even have to be the same account. I have my work account control my my personal account and vice versa. Um and so, that's why he's hiding the QR code cuz we could actually control it. We don't have to be on his account. Okay, let's keep going with how to do it. >> Simply take out your phone, point it at it, and it'll open up ChatGPT and ask you to accept the connection. And then from then on out in the ChatGPT app on your phone, you can connect to your desktop computer. So, then on your phone, there's this little remote tab. You just click that. And then at the top, you can see I'm >> By the way, you see he's showing that one computer, it's called Matthew Mac Studio 6 local. And if you look to the right of that, he has another one because he could from that phone control multiple Codexes on multiple computers. >> now connected to my Mac Studio. And all of the threads that I have on my Mac Studio are now available to me from my phone. I can continue coding from anywhere. >> I like that he is a continue coding from anywhere guy and a continue building even when I'm stepping away guy. He's big on talking about goals. I wasn't sure if I should include it, but I think when you've got a big project, this will be really helpful. So, I want you to see this. >> I'm going to show you something incredibly useful {slash} goal. {slash} goal makes chat GPT run for a very long time. But, not just for the sake of running a long time, it is for specifically achieving some goal. And you're basically telling chat GPT continue to run until you hit this goal. And a goal can be one of two things. It can be something verifiable. Continue until the speed of my website is 50% faster than it is today. Or, it can be LLM as a judge. And you can say, "Continue until the speed of my website is as fast as you think possible." And so, you can see the difference. One is very verifiable. It is something that is a concrete number. And the other you're leaving up to the LLM to decide. But, either way you can do that. So, as soon as I hit {slash} goal, you can see this little goal right here. And then, I will describe the goal and define a measurable outcome. And once I do that, this can run for a long >> Yeah, he's had it run for days and days. I've never needed that and I don't want to burn through credits that much or through tokens. So, what I do is I say, "{slash} goal, here's the outcome I'm looking for." For example, I want to match my new site to an old Squarespace site that I had and don't stop until you've checked and they've gone uh and and it looks exactly the same. Or until you've hit five rounds and then stop and I could evaluate it. I don't do what he does. He's had his go I think for as long as 14 days. That's way too much for me. Okay, speaking of not burning through tokens, let's talk about model choice. >> fixed quota that they give you each week and you don't want to just blow through the quota using kind of the best biggest model on a simple task. So, right here in the chat you click right there and first you select the model. You have 5.6 and three flavors of that, Luna, Terra, and Soul. You still have 5.5, 5.4, 5.4 mini. I'm mainly using these three. Luna and Soul are my go-to. I'm either using Luna to do things really quickly and save money or I'm using Soul for my most complex task. Now, within Soul, once you selected that, you also select your thinking effort. For me, I'm usually using medium, high, or extra high. I rarely use light and I rarely use the max setting for Soul. And the way that I think about whether to use Soul or Luna is obviously based on the complexity of the task. For the really hard task, I'm using the bigger model Soul. So, let's say I'm building a website. I will generally use Soul to actually build the website and then the fine-tuning of the website, like changing fonts or colors or moving text around, moving images around, that can all be done with Luna because it's capable enough for those simple changes. >> I have to admit, I don't do that and I'm constantly running out of uh tokens and I'm constantly adding more and paying for credits. It's a real pain in the butt. Um but I I I just keep thinking I'm I'm working. I don't want to have to keep switching models and models and models, but maybe I need to do that more often. Let's keep going here. I want you to see how he has a really big impressive request. Check this out. >> Review the latest sales and inventory data, then come up with one special to run over the weekend based on the data. Create a marketing idea to promote it on Instagram, generate the image in our style based on @ logo, suggest a caption, then create an @ derive folder and save the assets there, then @ Gmail the link to me. So, this will use two different plugins in one task. >> And the @ is how you reference plugins. I have not used the reference at times and sometimes it'll figure out what the plugin is and sometimes it'll go it'll go nuts. Um but I think it is important to say which one. >> And this will use a pre-built skill for image creation, which utilizes GPT image two, the best image model for this type of thing right now. It's all finished and the image looks awesome. I'll read the markdown file first. It went with the latency lemonade and programming pretzel. It has the reasoning below, how many pretzels are on hand with 26 above par that need to be used by Monday. There's been positive customer feedback on this combo, so a bunch of solid reasons for this choice. It lays out the Instagram campaign, then a suggested caption that looks solid. And the image looks amazing. And I can go to Google Drive and it created a brand new folder, then saved the files there. So, now I could share that with the person in charge of socials. And then for the email, it says my Gmail isn't connected. Um honestly, I never really use that in here, so that's probably true. I can go reconnect it and say send it. And actually, it's still not able to use the Gmail connector. I need to check the permissions there. I've had it draft emails before, so I'll see what's going on with that plugin later, but it's actually deciding on a workaround. It says it's going to open up the browser and see if I'm logged in there, which it looks like I am. I logged into this profile here for something else in the past, so this is actually ChatGPT controlling this browser, not me. The glowing one is the chat GPT cursor. So, it found me, typed out the full email with the drive link, and then it's clicking send. And I guess I can actually see right here that the email went through. I was just sending it from this email to it >> to itself. This is a very interesting if use case. I specifically picked it because it has ad design, it had reasoning, it's connecting to multiple tools, it is creating files and putting in Google Drive. It screwed up or he screwed up by not having the connection to email. It pulled up the browser, which is very powerful within the new Codex chat GPT app, and then it started working. When we're talking about this becoming an agent, it is an agent now that has plugins, tools, a browser, desktop control as we'll see. Um really, really very powerful and um and again, you can use all this stuff remotely, you can schedule it. Let's keep going because I said you can uh operate it can operate your computer. Here's one simple example. >> I'll run my first prompt. Here's a messy folder from a Vibe Coding Cafe, the context window. Clean it up by organizing the files into folders, then create one spreadsheet that pulls everything together so I can see what's going on with the business. I'll send that off, and now it gets to work with that folder as its workspace. And here's what that folder looks like now that it's starting with, just a bunch of random files in here, CSVs, PDFs of invoices, sales logs, text files. Work is going to analyze all the different files in here. It can understand all of these and extract the information from them. I'll come back once this is organized and take a look. I'll pull this up now, and you can see from all those unorganized files, I now have 11 folders in here. Chat GPT created all of these on its own, and it moved the files around and organized them into their proper folders. And they are all categorized correctly. Then the output folder down here should have the full spreadsheet where it combined all the data within those files, and it looks great. This is a way better spreadsheet than I'd ever than he'd ever make. It is really incredible to see it doing it. I think this is a basic example and I didn't want to get carried away with all the different computer use examples. There are plenty of videos on that. I just wanted to show that an agent can do that. For me, one best example that I had was I used Hermes agent. I was in Iceland. I had problems with it not using the right models. I attached one and then a backup and then a backup and for some reason I guess I didn't do it right. I just went to Codex on my phone and I said, "I want you to control my computer, get into the Hermes app and then fix it. I want it in this order." Now, truthfully, it made a mistake the first time where it it said it did it and then I went back to Hermes and I said, "What model are you using?" And it was using the wrong one. I went back to Codex. I said, "Hermes told me you're using it's using the other one, not what you're supposed to do." Codex said, "I made a mistake." And then it went back in and it clicked around on the Hermes agent app and made it work with the way that I wanted it to. And that's the power of this type of agent. An agent that can control your computer and control it well. I see what they're what they're going for here. Let's look at one last creator that was Futurepedia. I want you to look at Riley Brown and he loves voice mode. He did a whole video on it. Um but I think it's important to include >> this little white circle button right here. And if you hover over it, it says start a new voice chat. Let's go ahead and do that. Hey Chat GPT work. Um how are you? >> I'm doing good. Thanks. How are things going? >> Yo. Um can you please open up it's in my notion. Uh there's a video database. It's called GPT work explained. Can you please open that up in the Codex browser uh the notion doc? >> Sure thing. I'll open that notion doc in the browser. >> No, you actually got the wrong one. It's just called GPT work explained. >> Got it. I'll switch to that one. Okay, I'm switching the Codex browser to the exact page GPT work explained. >> Yes, exact. All right, and you can do this both with voice on the computer or with voice on mobile, which is such a great way to interact with your agent. Um finally, [snorts] there's another way that I like that he did it. I don't know if you're Tell me if you're as if you're as excited about this one as I am. I find this to be a very helpful way to interact with an agent. >> Intro. Can you please write it in an editable text block? >> That's the thing I didn't even know you can do. An editable text block. >> Okay, so this is going to prepare the intro that it's writing in a text block that I can edit. And you can see here it's using the Notion doc to understand what the episode's about, and here it gives me a response in this little block, and I love it. It allows me to just focus directly on this section right here. I can full screen it if I just want to focus on it. I can very easily describe edits to this document, and if I want to edit it directly, I can just edit it directly, right? I can type into it hello, and I can kind of ask AI to make changes or I can make changes directly. And if we go >> That's really helpful for me because sometimes I want um Codex to create a rough draft of something, and then I want to edit it, but I don't want to copy it into Notion and then edit there. Um it's also very helpful if it's asking you a bunch of questions, ask it to put the questions in an editable uh text block, then you can write the answers underneath its questions, you know, basically fill it in within this, and then send it back and say, "Here, I've given you all the answers." It helps me to be able to do that. And one last thing about this is if you continue the conversation from the desktop to mobile, the editable block also shows up there. So, those are the the creators that I think were most helpful for me to understanding how to turn Codex or ChatGPT into an agent. Now that you know how to do that, I think you might really enjoy what everyone else is calling Jarvis mode, which is a voice mode, and so I've got a video here where I do the same thing for voice mode. Subscribe if you like this, and I'll see you in the next video.