How to Build Your Own AI Employee with GPT 6 Astra

0:00GPT-6 Astra is the most powerful model OpenAI has ever released, and the first thing I did with it was give it a bunch of work I would normally hire an employee for, and the result really surprised me. In this video, I'll show you exactly how to build your own AI employee with GPT-6 Astra, and by the end, you'll know whether it's worth your money. Every prompt I'm using in this video is in the description, so you can copy them straight in and swap in your own files. And for this, I'll be using the ChatGPT desktop app because it has the work tab at the top. If you've used

0:30work on Claude, this is a similar idea, but if you haven't, work is basically the side of ChatGPT where you give it an actual job, like a folder of files and a brief, and it goes off and does the whole thing by itself. So, the first thing I'll do is click the model picker in the prompt box and switch it to GPT-6 Astra. Next to it is the effort level, and that sets how much thinking it does before it starts. So, you can set it manually between lower, medium, and higher. Then, I'll click the project chip, choose new project, select a

0:57folder on my computer, and create it. And now, Astra can only see what's in that folder and nothing else on my machine. I keep a separate project for each type of job, so invoices go in one and contracts go in another, and that way the model never mixes files from two different jobs. The folder I'm giving it has this quarter's vendor invoices, and that's about 60 PDFs, plus the

1:17spreadsheet we use to track payments. So, in the prompt, I'll tell it to read every invoice in the folder and match each one to a row in the payment spreadsheet. Anything that's been paid twice, paid late, or not paid at all, I want that flagged. At the end, I want a new spreadsheet with the full match, plus a one-page memo I can forward to the relevant department. One thing I recommend is to write the brief like you'd actually brief a new hire. Give it what you want done, where the files are, and what the finished thing should look like, because if any of that is missing,

1:45it just comes back with a question instead of a result. And honestly, that came back really clean. It works through the folder on its own, and the files it makes open in a panel on the side, so I can read them without leaving the app. And it's reading the 60 invoice PDFs straight off, so I'm not the one keying numbers out of each document before it can even start. The spreadsheet has every invoice matched to a row, and the memo is one page in plain English that has those flagged invoices and the total at the top. And that's what Work actually does differently, because the normal chat can't even open a folder on

2:15my computer. So, it has nothing to audit. But in Work, it goes through every invoice and gives me files I can send straight to my team. The memo is only as good as the folder you give it, though. So, if your spreadsheet is a mess, then the memo is definitely going to be a mess as well. But all of that happened inside a single folder, and

2:31teams often work across several tools. So, that's what I'm going to test next. Managing leads is one of the most repetitive jobs on any team, and it's exactly the kind of work I'd hire someone for. There's no integration to set up for this, because Astra has computer use, which means it can open the browser, click through the pages, and type into the fields itself on literally any app you have open. And the reason I use Astra for this instead of just setting up an integration is because computer use works on whatever app I already have. So, if we switch CRMs next quarter, the same brief still

3:00works, and I don't have to rebuild a single thing. There's no API to connect and no fields to map, either. And if I can click through the CRM myself, then so can Astra. So, I've got a spreadsheet with 30 leads from a conference. I'll brief it to open our CRM in the browser and add each one as a new contact with the company, the title, and the note from the conference. Every one of them gets tagged with the event name, and any lead it can't place goes in a list for me instead of guessing. I recommend naming the columns in your spreadsheet the same way the CRM names its fields,

3:29so the model doesn't get confused and put things in the wrong field. And honestly, this one is really easy to run, because I don't even have to watch it. It works through the whole list in the background while I'm doing something else, so the chat doesn't show it clicking around. But when I open the CRM after that, the new contacts are already in there, and the last 10 are as clean as the first. On my machine, that took about 6 minutes for all 30 leads. It did stop once, because it was about to save a contact that already exists in the CRM

3:56under a slightly different company name. That's because the prompt tells it to pause on anything that looks like a duplicate and ask me before it carries on. So, I'll tell it to create them as a separate contact [music] and it picks the job back up from there. And that one came out really well. 28 of the 30 are in with the tag on them and the other two had no company or email. So, they're in a short list at the end for me to check. So, I have one rule for what's allowed to run without me. Can I undo it in 5 minutes? If it's inside my own

4:22tools and I can undo it, like data entry or a report, [music] I just let it run. But if it goes outside the company, like an email to a client or a payment, it gets everything ready and I have to press the button myself. And I write that into every brief so it stops and asks every single time. Both of those jobs weren't really that long though and the long jobs are where AI usually loses track. So, I'll give it one of those next. So, I'll try building a small tool that takes the sales export we get every Monday and turns it into the report the

4:50board gets. And for that, I'll go to the top left menu and switch from ChatGPT to Codex because Codex is actually the part of ChatGPT that lets you build actual apps inside it. It runs on the same GPT-6 Astra [music] and I'll point it at a project folder with two things in it, which are last month's sales export and the PDF of the report the board actually got. In the prompt, I'll ask it to build a tool that reads the export, produces the report in the same layout as the PDF with the same five charts and to test it on last month's file until the numbers

5:21match the PDF before it comes back to me. Now, it just starts writing the code and running it on last month's numbers. So, I'll close the window and leave it completely alone. And by the time I came back to it, it was done. And this held up really well. It built the whole tool and tested it against last month's report by itself and every number matched except a rounding difference on one chart. The one difference it found, it flagged for me rather than changing the number to match. And honestly, this opens up a lot of freedom for me because whenever I need something like this built now, I can just do it from here in

5:51an afternoon instead of hiring someone for it. And that goes for pretty much any small tool a team needs. It's not only the tools I'd have built myself either because if there's some small app I pay for every month, there's a good chance I can have this built at once and stop paying for it. That's one less subscription to pay and I don't need to know how to code any of it because Astra writes the code, runs it, and checks its own work before it ever comes back to me. Now, the [music] context window

6:15actually plays a really big role here. And if you don't know what that is, it's basically how much of a conversation the model can keep track of at once. A long job like this used to run past that. So, halfway through the model would forget what it decided at the start and you'd get something that half worked. But

6:30Astra keeps its own notes as it works. So, everything carries over from one window to the next and it can search back through the earlier part of the job whenever it needs something from an hour ago. So, when you're briefing a long job, give it the finished example and something to check its work against like I did with that PDF. Because the example tells it exactly what you want. But there are certain jobs that need to get done without me having to tell it every single time and that's what the next feature solves and it's called scheduled tasks. In Work, you can set a task to

6:58run on a schedule or to run when something happens in an app you've connected. A report due every Monday or a contract that needs filing used to sit on my list until I remembered it. But now it just happens on its own. So, the first one I'll set is the Monday report and I'll tell it to run the tool from before. Every Monday at 7:00 in the morning on whatever export landed in the folder. So, it's in my inbox before I'm even at my desk. Then a second one watches Gmail and I'll tell it that when a signed contract comes in from a client, it saves the PDF into the

7:26client's folder, adds a row to the contracts tracker with the value and the start date, and drafts a two-line welcome email for me to send. The first contract that came in got filed in about a minute and by the time I opened my inbox, the row was already in the tracker and the draft was waiting for me. And that's when it starts running like an actual part of the team, because I didn't open the app, nor did I write the brief. The contract came in, and the work just got done. The last step is a draft I send myself, because that's the same rule from earlier, and it doesn't

7:53change just because nobody's watching. While those two are running, I can check on any task from the app, answer a question it has, or change direction, and it carries on from there. And between meetings, I hand off most of my work through the voice button on the desktop app, because in Work or Codex, you can just talk to it. So, when I say pull up the Q3 pipeline numbers and draft me a summary for the board, it starts that as a task of its own, while I keep talking to it about something else. So, the pipeline summary is being written in one task, while I'm dictating

8:21the agenda for the next call, and neither one is waiting on the other. So, is it worth your money? For me, yes, because it's already doing work I'd otherwise hire someone for. [music] And that pause on the duplicate contact is the reason it gets more of my work every week, [music] because when it isn't sure, it asks me instead of guessing. And once one of these jobs is reliable, the next step is connecting a few of them into a system that runs a whole department. So, watch [music] the video on screen now, where I build exactly that. Thank you for watching,

8:49and I'll see you in the next one.