A 32-minute walkthrough turning ChatGPT's desktop app into a full AI staff: persistent memory, recordable skills, connected real tools, and specialist workers that execute jobs instead of just answering questions.
Running a business on GPT-6 Astra means replacing the chat window with a system that has persistent memory, reusable skills, and real tool access, so a team of AI workers executes recurring work while a person only reviews the exceptions.
Who This Is For
Read if. Skip if.
READ IF YOU ARE…
A business owner who wants to move recurring email, document, CRM, and reporting work off a person's plate and onto a system the business actually owns.
Someone already using ChatGPT only as a chat window who wants to understand what memory, skills, and connected tools add on top of that.
An operator willing to pay for a subscription plan and spend real setup time writing or recording skills and connecting tools, not just chatting.
SKIP IF…
You want a fully hands-off system that runs the company with zero human review; the video argues directly against that.
You need an enterprise security or compliance evaluation before connecting AI to production tools; this is a practical walkthrough, not an audit.
TL;DR
The full version, fast.
A normal ChatGPT tab forgets you, can't touch your tools, and only works while you're typing, so this guide rebuilds it into a system with four layers: memory (Projects and an agents.md file hold the business's context permanently), skills (recorded or written playbooks that trigger automatically), connected real tools (email, CRM, drive, and calendar via OAuth), and specialist AI workers that execute goal-based tasks in the Work tab instead of just chatting. Control comes from plan mode, staged approvals, and permission tiers, not from trusting the system blindly. The advice is to start with one workflow, get it working three times in a row, then expand, since heavy usage runs $100 to $200 a month per person and messy source files produce messy output.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Cold open stating the promise: a complete guide to running a business on GPT-6 Astra, framed against businesses still paying people to do work AI could finish in seconds.
01:40 – 03:07
02 · What this actually is
Reframes the setup as one place holding the whole business plus a team of AI workers with real tool access, with the operator in the middle, versus just using AI in a browser tab.
03:07 – 03:34
03 · 3 problems with a normal chat window
States the three failures of a normal AI chat: it forgets you when the tab closes, it can't touch real tools, and it only works while you're typing.
03:34 – 04:28
04 · Setup: the desktop app
Introduces the free ChatGPT desktop app and its three tabs: Chat, Work, and Codex.
04:28 – 05:47
05 · Which plan to be on
Covers the three subscription tiers and recommends starting on the flat-price Plus plan rather than usage-based billing.
05:47 – 08:07
06 · Memory: projects
Explains Projects as the workspace that holds a business's context and files, plus Astra's ability to keep its own notes across long jobs.
08:07 – 09:13
07 · Custom instructions and agents.md
Distinguishes account-level personalization instructions (Codex-only) from the agents.md file read automatically every session.
09:13 – 11:38
08 · Skills: your saved playbooks
Defines a skill as a saved standard operating procedure that the system recognizes and follows automatically when a matching request comes in.
11:38 – 12:17
09 · Running a skill on a real sponsor email
Demonstrates the sponsor-reply skill firing automatically on a real inbox request.
12:17 – 13:21
10 · Making a skill with the skill creator
Shows the built-in skill-creator flow turning a plain-English description into a reusable invoicing skill.
13:21 – 16:09
11 · Recording a skill instead of writing one
Demonstrates the screen-recording skill builder: do the task once on screen, the system watches and writes the skill.
16:09 – 19:26
12 · Connecting your real tools
Walks through the plugin marketplace, connecting Gmail, Notion, Slack and other tools via OAuth so the system can act inside them.
19:26 – 19:59
13 · Why fewer plugins is better
Warns that connecting every available plugin adds heavy token overhead versus connecting only the few tools actually used.
19:59 – 20:28
14 · The work tab and your AI workers
Explains how the Work tab takes a goal, breaks it into steps, and hands back a finished deliverable instead of a quick answer.
20:28 – 21:37
15 · Plan mode: a proposal for an HVAC company
Live demo building a one-page proposal and spreadsheet for a dummy HVAC client using plan mode.
21:37 – 23:09
16 · Why specialists beat one big worker
Argues for narrow specialist workers, each in its own project, over one general worker that gets confused with too much context, and gives OpenAI's recommended build order.
23:09 – 23:46
17 · How to stay in control
Lays out the three control levers: plan mode, approvals, and permission levels.
23:46 – 25:33
18 · Approval settings
Details the ask-for-approval, approve-for-me, and full-access permission tiers, and the rule that money and client-facing sends still need a human sign-off.
25:33 – 27:01
19 · Scheduled tasks
Shows setting up a recurring task in plain English, covering title, schedule, device, and notification settings.
27:01 – 27:44
20 · Local vs cloud tasks
Contrasts local tasks (device must be on) with cloud tasks (run anytime, can trigger on events like a new email).
27:44 – 28:27
21 · Using it from your phone
Shows pairing a phone and laptop to the same account for working from anywhere.
28:27 – 29:24
22 · Computer use
Explains computer use as direct screen control for tools with no API, such as old CRMs or plain web forms.
29:24 – 30:02
23 · Where it falls short
Cautions that OpenAI's adoption numbers are self-reported and describe well-defined jobs, not a system left to run the whole company unsupervised.
30:02 – 30:39
24 · Cost and messy data
Covers real-world monthly cost ranges for heavy users and the garbage-in-garbage-out dependency on clean source data.
30:39 – 31:04
25 · The golden rule: one workflow first
States the adoption rule: pick one workflow, get it working three times in a row, then expand.
31:04 – 31:55
26 · From doing to deciding
Frames the real change as a shift in the operator's role, from grinding through tasks to directing and checking work.
31:55 – 32:37
27 · The 3 leaks in every business
Closes with the three recurring time leaks every business has: chasing people, manual information moves, and unread reports, plus the closing offer.
Atomic Insights
Lines worth screenshotting.
OpenAI reports that 98% of its own staff use ChatGPT's agentic Work tab daily, and about 90% of the work inside its legal, finance, and recruiting teams runs through it.
A normal AI chat window fails a business in three specific ways: it forgets everything when you close the tab, it can't touch your real tools, and it only works while you're actively typing.
Turning a repetitive task into a reusable skill just means writing the steps down once, since the system then recognizes matching requests and follows those exact steps automatically.
A screen-recording skill builder lets you demonstrate a task once on your own screen, and the system writes the reusable instructions from watching you do it.
Connecting every available plugin instead of just the two to four tools actually used added roughly 300,000 tokens of context overhead to a single job before it even started.
Splitting AI work across narrow, single-purpose specialist workers avoids the context bloat and fabricated answers that come from handing one general worker too much information at once.
Plan mode forces the system to write out its intended steps and wait for explicit approval before touching anything, which is the main lever for staying in control of an agent with real tool access.
A cloud-based scheduled task keeps running even when the laptop is off and can trigger on events like a new lead email, while a local task only fires while the device and app are both open.
OpenAI's own guidance is to build rules first, then skills, then connect tools, and only add specialist workers once a task is clearly separate enough to need its own lane, not to launch twenty workers on day one.
Heavy users of agentic AI work typically land in the $100 to $200 a month per person range, not because the tool is broken, but because that is what actually running a team through it costs.
Every business tends to leak time in the same three places: chasing people for approvals or replies, manually moving information between systems, and maintaining reports nobody reads until they're wrong.
The reliable path to adoption is mastering one workflow until it works three times in a row before expanding to adjacent processes, rather than automating everything at once and abandoning it when it gets messy.
Messy, disorganized source files and notes produce messy AI output just as reliably as clean, organized input produces clean output.
The real shift this setup produces isn't a faster tool, it's a change in the operator's job: from personally doing the tasks to directing and checking the work of a team that does them.
Takeaway
How to turn ChatGPT into a working team, not a chat window
AI OPERATIONS SETUP
Running a business on GPT-6 Astra comes down to four layers stacked in order: persistent memory, reusable skills, connected real tools, and specialist workers, each one added only once the layer before it is solid.
02What this actually is
The core shift is from using AI, people opening a tab to ask questions, to being AI native, recurring work running on a system the business owns with a person only touching exceptions.
Picture yourself as the operator in the middle, with a team of AI workers on one side and your real tools on the other, rather than a single chat window you talk to directly.
033 problems with a normal chat window
A standard AI chat window has three structural limits: it forgets everything once the tab closes, it can't reach your real tools, and it only works while someone is actively typing into it.
Any system meant to replace a chat window needs to solve all three limits at once, since fixing memory alone or tool access alone still leaves the other gaps open.
04Setup: the desktop app
The Chat tab is for normal back-and-forth conversation, the Work tab takes a goal and returns a finished deliverable, and the Codex tab is for coding tasks, all inside one free desktop app.
05Which plan to be on
A flat-price subscription plan is worth starting on instead of pay-per-use pricing, since usage-based billing on agentic tasks creates unpredictable costs.
Start on the cheapest paid tier that unlocks agentic features and only move up once that tier is obviously not enough, rather than guessing upfront.
06Memory: projects
A project is a persistent workspace that holds a business's ongoing context, source files, and instructions, so that context doesn't have to be re-explained in every new chat.
On long jobs the model writes itself notes across the task instead of compressing everything into a summary and losing detail, which is functionally like onboarding a new hire once and having them actually retain it.
07Custom instructions and agents.md
Account-level personalization instructions only affect the coding-focused tab, while a separate agents.md text file living on the computer gets read automatically every session and can hold standing business rules and structure.
08Skills: your saved playbooks
A skill is a written-down standard operating procedure that the system recognizes and follows automatically when a new request matches it, without needing to be looked up or invoked by name.
Writing a skill can be as simple as a short description plus a few concrete instructions, what to read, what rule to follow, where to save the output, rather than an exhaustive script.
A dedicated skill-creation flow can turn a plain-English description of a recurring job, plus example files, into a reusable skill without hand-writing the instructions.
11Recording a skill instead of writing one
Any repetitive task done in a browser or app can become a skill by doing it once on screen with a recorder running; the system writes the instructions from watching.
Skill recording is meant for stable, repeatable processes with a clear finish line, not open-ended work like an entire sales process attempted in one long recording.
12Connecting your real tools
Real tools connect through an OAuth login or API key from a marketplace of over a thousand pre-built integrations, turning the system from something that talks about the business into something that acts inside it.
Connecting every available integration instead of only the handful actually used adds heavy, unnecessary context overhead to every job, so it's better to connect only the tools that are part of the daily workflow.
14The work tab and your AI workers
The Work tab differs from a chat by taking a goal, breaking it into steps, working through them, and handing back a finished deliverable, staying on a task as long as it takes.
Splitting work across narrow single-purpose workers, each its own project with only the context it needs, avoids the confusion and fabricated answers that show up when one general worker is overloaded.
The recommended build order is rules first, then skills, then connected tools, and only then specialist workers added one at a time as a genuinely separate need appears, not all launched together on day one.
17How to stay in control
Plan mode makes the system write out exactly what it intends to do before touching anything, so nothing happens until that plan is explicitly approved or corrected.
An approval setting sits between asking permission for every single action and giving completely unrestricted access, defaulting to it means only genuinely risky actions require a stop-and-ask.
Anything involving money leaving the business or something going out to a client should still get a human sign-off regardless of how much autonomy the system is otherwise given.
19Scheduled tasks
A task can be set up by describing it in plain English rather than configuring every field, with the system asking follow-up questions to fill in the missing details.
A task scheduled to run locally can see the computer's files but only fires while that device and app are on, while a cloud-scheduled task keeps running when the laptop is off and can trigger on events like an incoming lead email.
Pairing a phone and a laptop to the same account keeps chats, recent work, and scheduled tasks in sync across devices, which matters for anyone not always sitting at one desk.
22Computer use
For tools with no API or plugin available, old CRMs, web forms, sites without integrations, a computer-use mode can take over the screen directly and click, type, and browse the way a person would.
Adoption statistics like '98% of staff use this' are self-reported and describe people running specific, well-defined jobs through the system, not the system running an entire company unsupervised.
24Cost and messy data
Agentic work tasks are metered by job size and cost meaningfully more under the hood than a quick chat reply, so heavy users commonly land in the $100 to $200 a month per person range.
Output quality is bottlenecked by input quality: disorganized files and chaotic notes produce a messy result just as reliably as clean, organized source material produces a clean one.
25The golden rule: one workflow first
The reliable adoption path is picking one workflow, getting it working correctly three times in a row, and only then expanding into adjacent processes, rather than automating everything at once.
Attempting to automate every process at once is the common failure mode: the team stops using it and the owner concludes the technology doesn't work, when the real issue was starting too broad.
The actual change this setup produces in a job isn't speed, it's a shift from personally grinding through tasks to directing the work and reviewing the outputs.
27The 3 leaks in every business
Every business tends to leak time in the same three places: chasing people for documents, approvals, or replies; manually moving information between systems; and maintaining reports nobody reads until they're wrong.
Owners most often automate whichever task personally annoys them the most rather than the one actually costing the most time or money, so it's worth identifying the real leak before building around it.
Glossary
Terms worth knowing.
Work tab
The ChatGPT desktop mode that takes a stated goal and works through multi-step tasks autonomously, returning a finished deliverable, as opposed to the normal Chat tab's single back-and-forth answers.
Astra
The GPT-6 model powering the Work tab's agentic tasks, referenced by name throughout as the engine behind the setup.
Project
A persistent workspace inside ChatGPT that holds a business's ongoing context, source files, and memory so it doesn't need to be re-explained in every new chat.
agents.md
A plain text file kept on the user's own computer that Codex reads automatically every session, holding standing business rules, offers, and team structure.
Skill
A saved, reusable set of instructions for a specific recurring task, created either by writing it out manually or by recording the task once on screen.
Plan mode
A setting that makes the AI write out its intended steps and wait for explicit approval before executing anything.
Computer use
A mode where the AI directly controls the screen, clicking, typing, and browsing to operate tools and websites that have no API or plugin connection.
“If you feed one worker too much information, too much context, it gets confused, it starts making things up.”
clear rationale for specialist workers→ TikTok hook↗ Tweet quote
23:20
“The goal isn't to turn your business into AI or a robot and walk away. It's just put a capable team in front of you with you checking the stuff that actually matters.”
thesis for the control section→ IG reel cold open↗ Tweet quote
30:10
“Heavy users landed about $100 to $200 a month per person. And that's not it being broken. That's just what running a team actually costs.”
“You are going to stop being the person grinding through the tasks and you become the person who is directing the work and checking the outputs. You move from doing to deciding.”
the video's closing thesis in one line→ TikTok hook↗ Tweet quote
The Script
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
17px
metaphoranalogy
This is going to be the complete guide to running your entire business on GPT -6 Astra. I'll go from start to finish, the exact setup that we use, and the ones that we deploy for our clients. Now with that said, just like most of you, you're using a normal AI chat through your browser.
And that has three different problems that makes it completely useless for running any real work in this setup. It fixes all three. And right now, most businesses, they are still paying to do work that AI could be finishing in seconds.
And the ones who have figured that out months ago, they are quietly pulling ahead of everybody. still doing this manually now most people they hear chat gpt and they picture a window something that you type into and copy the answers out of now not only are they wrong about that but that one assumption is literally costing them hundreds of thousands if not millions of dollars now i've moved my entire team and my business over to this setup and i've never written a single line of code and this was never just about chatting it's about having a team of ai workers that actually do work for you instead of just talking about it now to add that openai the company that actually makes this 98 of their own staff they now use this every single day and about 90 of the work inside of their legal their finance and recruiting teams runs through this as well now just a quick intro if you don't know me already my name is nick i've spent the last few years implementing ai systems inside of real businesses and teaching thousands of entrepreneurs how to actually sell ai effectively everything that i'm going to be showing you today like the project instructions the skills the recorded workflows all of it i'm putting it inside of my completely free
school community, along with plenty of other resources to help you grow your AI business to over $10 ,000 a month. Link will be down below in the description, but enough said, let's dive into it. Before we touch anything at all, let me just reframe what this thing even is.
Now, the simplest way to think about this is you're going to have one place that is going to hold everything about your business, plus a team of AI workers that can actually do the tasks and use your real tools. So you're not going to be simply chatting with it. You're handing it your entire jobs.
And here's the whole idea. just a single picture so this is you in the middle right here you're effectively the operator now on one side this is going to be your ai team on the other side this is going to be all of your real tools.
So this is going to consist of different things like your email, your documents, your files, your CRM, any stuff like that. And you are just going to be sitting in the middle and you point to different things. And this is the difference between using AI and actually being AI native.
Using AI, this just means your people, they open up a tab. AI native, this means that the recurring work, it runs on a system that the business actually owns, loaded with how you actually do things, how you conduct all of your process and a person. they only touch the exceptions.
And that's what we're going to be building today. OpenAI, they actually put out a report that I just mentioned earlier in the intro where their legal team, their finance team, recruiting team, people who have never written a line of code before are running most of their work through this as well. And out in the real world, non -developers are now the fastest growing group using this, something like one in five users using this, doing research, writing any spreadsheets, conducting any presentations, building out any contracts.
just by describing what they want in plain English. And this is where those three problems with a normal chat window are coming in. So number one, it forgets you the second that you close the tab.
Number two, it cannot touch any of your real tools. And number three, it only works while you are sitting there typing at it. Now, every piece that I'm about to show you kills one of those and I'll point at each one as we go on.
So that's the entire shift here. It's not just a smarter search box. It's a place to actually run your business with a team.
But anyways, let's set it up and it's gonna be a lot easier. easier than you would actually expect jumping into the setup this is the part that most people are actually scared of and it's honestly the easiest part because everything I'm about to show you it lives in this one desktop application this is the chat GPT app that you see down below but you have a couple of different tabs at the top we have chat and we also have work now chat this is just where the normal chat window everybody is used to they're familiar with work this is the one built for exactly what we're going to be doing where you give it a goal and it comes back with the finished thing and then we also have if we flip up to the top we have codex now codex my favorite this is the coding one and we're going to be dipping into it just once for one file and that's it so first go ahead download the desktop application from openai's site if you have not already it's completely free you install it like any other app that you have before when you open it of course just sign in with your chat gpt account click on the work tab because we're going to be starting from there now quick tip as obvious as it may seem or not you want to be on these
subscription not paying per use through their developer platform now the reason why is because with the subscription you're going to be just paying one flat price as opposed to getting all of these unexpected costs rising and thrown at you every single month i'll just cover the plans real quick the app is going to be on every plan even free but astra and the work tab it does need a paid one so Plus is going to be about $20 a month.
And that's where I would start pretty much everybody, including you watching this video right now. Above that, there's going to be pro at about $100 a month and a bigger one at $200 for people who live inside this all day and are just going through completely blowing their usage. Now, one thing worth knowing, and OpenAI does say this themselves, is that heavy users tend to land somewhere in the $100 and $200 a month range.
So start on $20 and move up when it's obviously worth it. Just really quick, I just wanted to mention that if you are a business owner and you are looking to become AI native, transform your operations, beat your competitors before they beat you and you have to play catch up.
and just increase your bottom line there's going to be a link down below in the description it's just a brief form that you have to fill out it's going to show you where to be starting the processes to be automating what you're actually losing right now by not implementing ai and what you can be getting back in terms of a return so it takes about 30 seconds to fill out you can get it for completely free link is down below in the description the first piece and honestly the most important one is going to be the memory and this is what makes everything else work because it is the fundamental layer that is going to be applying to pretty much everything else that you are going to be using.
Down below our little chat box, we have a thing called projects. And the easiest way to think about it is a project is where your business lives. You can give it a name.
So you can see I have some different ones like how to get clients, our three -day challenge, some things for email. uh creating short form slides all that different stuff now to actually create one if you have not already you can just click on new project we can call this youtube demo and we can just add in any of our previous files we want to provide as much context to what this project is going to be about we'll also be able to soon add some instructions but here you can give it some default memory and to elaborate on this a little bit Default memory, this is just going to be able to allow access from outside chats and vice versa.
So if you want to interact with something through this window right here, and then you want to go to another one and be able to access this project's memory, you'll be able to do so. Vice versa, you can just have it be project -only memory. Within your source files, if you do have any SOPs or any outstanding documents or files that can give ChatGPT some context about who you serve, how you talk, and your prices, your rules, the way that you like things done.
That's where you're going to be wanting to provide this. Let's go ahead and create this project. We will add the project context from YouTube demo.
Go ahead and say yes to that. Now, one of the models, Astra, they actually have recently added something that I really didn't expect. And that's just on any long jobs, it's going to keep notes for itself across the whole thing.
So when a test does run past the point where an older model would have started to forget the beginning, Astra would just write itself some notes and goes back to them instead of just squashing everything into a summary and losing it. And the way that I like to describe that is you're just onboarding it once like a new hire and it actually remembers the onboarding.
for you to actually provide your project instructions you can simply just chat in this window give it some breakdowns you know for example if we are creating a project that is going to be for our onboarding process we may just break down what the process may look like how we actually speak we can give it some examples and references to anchor from so maybe that's going to be some emails or maybe our click up form that we send to all of our clients who are starting to be onboarded beyond that you can actually go into your account settings if you navigate to personalization you can add your instructions here now this right here is only going to be relevant to codex so if you're going to be using this quite heavily this is where i would put your custom instructions here now one thing worth knowing there's actually going to be a deeper version of this if you are going to be utilizing the codex side of things and that's just going to be a plain text file it's called agents .md this is what it looks like right here it's just going to be a folder that sits on your computer
and it gets read automatically every single time. So you can see what this actually looks like, how we work, what we do, who we serve, the offers, how we talk, some hard rules that needs to be following, where things live, the team, and how we are actually structured, how I like to work, all that sorts of stuff. And the most important things, if we are to templatize this a little bit, we're going to have all of the big rules at the top, some specific notes inside of specific client folders.
That's how I run it for all the technical stuff. And there's a command called slash I -N -I -T. So I run it for all the technical stuff.
If you do not already have something like this constructed, then I'll just go ahead and pull up ChatGPT, start interacting with it and having it come up with your simple agents .mt file. That's the memory. Let's go ahead and give it some skills.
So to explain this part extremely simply, a skill, it's just a saved playbook. It is more or less a standard operating procedure. So it's how you teach something, how you teach your AI more specifically, how to do a specific job in your format in the way that you normally would.
So it does the same way that you would be doing it. every time after that now if you've ever written any sort of sop that is exactly what this is except it actually gets followed step by step if i type slash and let's just put something in you can see all the example skills right here so this one it's going to be building secure sandboxed code executions this one actually helps you create skills there's so many and we'll touch on more of these in just a second but the magic with this is you don't have to dig these up and run them it notices when a job matches one of your skills and pulls it up on its own so when i say something thing like draft or proposal, it's actually going to go off and see that I already have a skill for that.
And it follows my exact steps in what the skill actually is comprised of in all of its instructions. So where they actually live, if we navigate to the sidebar, go into the plugins, and we can see the skills up at the top right here. If we expand this just a little bit, you can see all the different skills that I have right here.
So some of them are going to be client reports, building out invoices, meeting briefs, creating proposals, replying to some sponsors. helping create an outline of our videos our newsletters creating linkedin posts the list goes on and on even for creating some youtube thumbnails every one of these different skills it was simply just a job that i was doing over and over and over again and i just wrote it down once and let me just open one up so you can actually see how simple it is so let's go to the youtube thumbnail you can see we have a brief description well i hardly consider this brief But we have a description up at the top, and then we have some more specific instructions down below.
You can see how expansive this actually is. This one is my sponsor reply skill. So this is going to be for drafting and replying to a named sponsor in our inbox.
It is grounded in the rate card and to save the draft in a specific folder. So the instructions, just read the named email, follow the workspace sponsorship rules, draft a short friendly reply. This is only a draft only, and then save the output.
So that is simple as it gets. If I wanted. to.
I could expand on this, how I want the emails to actually sound, how I typically speak. I can give it some examples of how I typically talk, how short I want them to be and what to always include, what to never touch without checking with me first. And that's the whole thing about these skills.
And now if I'm just going to go back and ask this, if we actually navigate into the ChatGPT work section, we can just go ahead and just ask it like, hey, I need you to create a draft or use the sponsor email skill. to go ahead and reply to my latest sponsor i'm actually going to flip from gpt6 astro because we don't need to use a crazy model for something this simple And we'll just run this off.
It's completely unnecessary for me to even say, use the sponsor email skill. It's always going to recognize that it has a skill for that. And just like that, about a minute later, draft created in the latest sponsor thread.
It asks for the agreement, so on and so forth. So I'm actually gonna go ahead and just say, delete that draft now. Now, if writing any of these skills feels like a lot to you, you can simply just click on the plus icon.
There's a record a skill, which we will get into in about a minute from now. But also we can simply just find where that skill creation was. Let's see if we can search it.
Skill creator. So this is going to create or update a skill. So now I can simply just say something like, I need you to go ahead and create a skill to automatically write my invoices.
I'm going to be providing you some files on what our invoices look like. And it's up to you to templatize that and try to do that moving forward for us. Whenever I ask you to create an invoice, I will be providing you with the amount and who it's going to be billed to.
And then we'll just simply have to run it and it'll create it for you. Additionally, if your team has already been using this or you guys just want to be sharing things across your work environment, then you can just simply download their skills. All it consists of is just a simple file on your computer.
they can share it to you. Or if you're sharing the same account, then you can just use that as well. However, it is incredibly easy to share them across teams, share them with whoever else.
Like I mentioned, there's actually another way to be creating your skills, and that's just recording them. And the way that it works is you do the task once on your screen, the normal way that you always would, and it watches. Now, when you stop, it writes the skill for you.
So if you think about what that means, every repetitive thing that you do in a browser or in an app, whatever it may be, hypothetically let's say there's a report that you export every single monday or the invoice that you file or the post that you put in your community the thing in your crm that you click through perhaps 15 times a week you don't write any instructions for it you do it once with the recorder on so let me stop talking and actually record one simply all we have to do click on this plus icon once again record a skill and it's going to automatically input this in now i already have it downloaded you'll just simply have to click on the button that says install and start setting up i can give it a little bit more of an expansive description inside of this however let's just run with this let's go to gpt6 astra and we will go to the light mode and we're going to run this off right now it's going to prompt us if we are allowed to use chat gpt to record the actions on our computer so you can see now it is automatically recording it so let's go into our go high level and let's just try to pull some of our booked calls we're going to go into our calendars
and just flip on the necessary closers for our teams, which would be myself and a couple other people. So here I can just click on each one and we're going to get the information from this. However, I also want to be, once I do exhaust this and go through all of the meetings from today, we're going to go into our opportunities section.
So for this, it's going to be just for B2B. Now we have a couple other pipelines, but it's going to be to be. And we're going to go to anything that came out after this.
So strategy call attended is what we're actually looking for here. So we click on any one of these, go through our process how we normally would. And we can just maybe show clicking around, copying all the information.
So we'll click on copy number, whatever it may be, and do the same thing for everything else. now to make anything more clear to the system we may want to provide some sort of brief description like i mentioned earlier but it's this simple so you just go through your process record on your screen and you give it to chat gpt so let's finish this now two things to know this is only available on mac right now and it does need a setting called computer use switched on so how you actually do that you just go into your settings and you can either just go to computer use if you can't find it you can just search everything up at the top computer use and just make sure you enable whatever necessary on the other note this is only built for stable, repeatable processes with a clear finish line.
So something that you are going to be doing every single time. I would not try to record your whole sales process in one go. But any case, that is the skills.
You either write them or you show it once. Now it's connected to our real tools. So far, everything has been inside of this app.
Now to actually run a business, of course, it needs to reach into the tools that you're already using. So your emails, your documents, your customer records, your CRM, and that's what this piece is going to be doing. So we go back into our plugins and this first side right here, this is where you can access, you know, over 1400 of these different applications.
So we have different ones like Microsoft Power BI, AWS Data Analytics, but the more important ones probably relevant to you is Slack, Google Calendar, Gmail, Google Drive. We also have Microsoft inside of here, if I can spell properly. We're going to access our Microsoft Power BI, our Teams accounts, our spreadsheets, Outlook email, all of the most important things that are effectively running your entire business.
You think of each one of these just simply as a plug. So for us, we have Slack. should already have this installed but all we have to do is just connect our accounts that we have built under our slack account and now it'll be able to read all my slack messages but on the flip side if we do it with email it'll be able to read all of my emails go into my google drive if that's what we are using inside of our operation and whatever it knows exactly where it's to be going just like i've shown you earlier is able to interact with my gmail actually create a reply for our sponsorship and get back to them so i was able to actually go in create that draft for us and if we would like it can automatically send it off but i get an idea of what i actually have installed myself if you go to manage and you can see i have cloudflare click up notion drive slack booking .com to you know actually book hotels i've used this to help me book some vacation trips pds documents i mean the list just goes on and on now the one that i've actually used probably the most of is going to be my meeting ai recorder where it's going to just go in
get all of my meeting transcripts. So for example, I've had plenty of team meetings where we discussed a lot of to -dos, a lot of action items that I need to be taking care of. So I take that transcript, I have it help, you know, come up with a plan for anything that I need to be enacting on the next step or any other action items.
And it's able to easily just look at every single line inside of that meeting, which is incredibly important for myself and my team's productivity. Not as recently, our team was just discussing a brand new campaign. So we need to do a lot of different things.
want to discuss our go -to -market plan there's a lot of things i have to be doing like setting up the funnels creating all the landing pages creating the copy so in my specific example i had it just go ahead and help me come up with a copy i provided all this different context to be able to get everything from the meeting what we discussed and all the important contacts to actually connect any one of these we can go to gmail and we simply just have to scroll down we click on connecting our account we can connect another one if we would like and we just have to use the oauth where we just log into our account it's that simple sometimes you'll have to provide an api key but it's very simple so we do the same thing for notion it just prompts us to log in with our gmail or whatever account our username and password and just like that it's connected now this is where you're going to start noticing that this isn't just talking about your business anymore the ai is going to be inside of your tools doing the work for you pulling requests pushing requests updating things this is where that slogan of you can now run your company with ai stops being a slogan because now one instruction
and it moves through several of your tools at once. Now, two things to keep in mind here. One, more is not better here.
So if you plug in the two, the three, four tools you actually live in and stop, that's what I would recommend. I measured this on my own setup and having every single plugin switched on added about 300 ,000 tokens of overhead to a job before it even started. And that's your usage just getting eaten for nothing.
And number two, just stick to the official well -maintained ones. I would use all of these trusted. plugs not some random thing that you found online so at this point your team should now have hands let's talk about the workers themselves okay so this one right here is going to be making this whole ai team idea actually seem real so first the difference between the work tab and a chat in a chat you can simply just be asking a question get a simple answer.
However, in the work section, you give it a goal and it breaks the goal into steps. It works through them and it hands you the finished thing. So if that's a deck, if that's a spreadsheet, if that's a document or a little webpage, whatever it may be, it'll stay on a job for hours if it has to just to accomplish the job.
So let me give it an example right here. So first, we're just going to turn on the plan mode right here. We're going to say, build me a one page proposal for Northside Heating, a made up HVAC company.
You can just use dummy data for this and put a summary into a spreadsheet i will run this off this actually goes to astra light because we don't need to go overkill with this first it's going to check the document and the spreadsheet guidance and then pin down the proposals focus and the format so it's listing all these different documents so reading the document skill reading the spreadsheet skill listing some of the files and you know just actually working the output now i'm using the document and spreadsheet skills to plan the proposal what should the proposal cover i'll say hvac installation so it will ask me some questions but more importantly it's writing the plan right here a plan mode this is how you actually stay in control so we can open this up here's north side heating proposal and spreadsheet here's the summary deliverables dummy project details and some verification now we can take a look at this and say okay this looks perfect or if there's some changes this is just where we have to tell jgpt
right on the left -hand side right here, or we can just say implement this plan. That simple. Now to take a step back and just look at some things at a higher level, this is going to be you at the top and under you are going to be your department.
So we have a proposal worker, we have a content worker, an operations worker, we even have a research worker. Each one, it's just simply a specialist that only knows its job. And the reason that we split them up instead of just building one giant worker that knows everything is because if you feed one worker, too much information too much context it gets confused it starts making things up and it gives you wrong answers it is inevitable however having specialists they do not have that problem they will not experience that context bloat so how do you actually get that to work inside of the application well you just run several work tasks at once each one inside of its own project so each one only has the context that it need if we were utilizing the codex tab there's going to be proper version of this where each worker is its own file with its own instructions and it hands work to them on its own that's what i use for all of my heavy stuff and pretty much my day -to -day stuff but for most businesses separate projects and separate tasks are going to get you where you need to be now one thing worth saying is that this is straight from opening eyes own guidance on how to actually build this out and that's not to go building 20 workers on day one
their order is to just build the rules first and then skills and then connect your tools and only then add your specialist workers so start with that one and you can add a specialist only when you've got work that's separate enough to need its own lane my personal take it's the same thing one good worker it beats a mess of 10 different ones that you never tuned a question that you probably are asking to yourself and you definitely should be is that if this thing can get into my files in my email whatever else it may be and actually do stuff on my behalf how do i stop it from doing something that i do not want it to well there's actually a few different things that you can do first is going to be the planning mode that i have shown off earlier now with this Obviously, you already saw it for the bigger jobs.
Say you ask it to clean up a messy folder full of client files. You don't want it diving in and moving 100 things around. So it writes out exactly what it's going to do.
You read it and nothing happens until you say go. And if the plan is wrong, you tell it what to change. Now, the second thing is going to be the approvals.
Now, out of the box, ChatGPT, it's going to be doing things very carefully. The second that it wants to send something out or change something that. really matters to you it stops and asks you first so it's like do you want me to send this and usually you're just saying back no let me read it first or yeah go ahead and you decide this one time or just trust it for the task so how you actually do this is you can just go to this little section right here you can ask for approval where it's always going to be asking to edit external files and using the internet or you can say approve for me where it's some sort of auto mode where it only asks for actions detected as potentially unsafe or you can give it full unrestricted access to the internet in any file on your computer which i would advise against to start with especially if you have important things on your computer or if you're working inside of a team actual business environment the one that i use daily is the approve for me it's my daily driver it does pretty much everything for me now of course i still have to approve some things every now and then but it's just some sort of auto mode and it's you know alternatively i would have to approve
every single thing that it wants to do if it wants to search something up if it wants to go into my crm and pull something i have to prove there i have to sit here approve everything which is obviously not very efficient the third thing is how much rope you actually give this thing if you do have some sort of compliance or security concerns just keep it on the careful setting that's what it's there for and the point of all this is that the goal isn't to turn your business into ai or robot and walk away it's just put a capable team in front of you with you checking the stuff that actually matters so anything with money going out or something getting sent to a client you still sign off that's not a limitation that's just how you actually run this properly so that is the control piece now let me just make it run without you babysitting it every second so now at this point you are still the one kicking everything off this next piece is simply about making the routine stuff happen on its
own and it's three different things the first one is going to be the scheduled tasks so it is going to be up at the top in our browser it should be pinned already for you now here if we want to create something go ahead and we simply just give it some sort of instruction so let's set up a scheduled task together first explain how schedule test works in chat gpt and then interview me to figure out what i need scheduled and when it should run so we're just using plain english here you can see as soon as you click on something it automatically creates this prompt for you so you don't really have to do anything except for just explaining in plain english you can talk to it whatever you want to do or you could also just set this up manually so we can go into this little edit section right here we can give it a title we can give it a brief description the details and what this will be running on whether that's your device or a cloud what this should be running in like what specific chat and the frequency that you want this to be running whether that's going to be daily weekdays or weekly and you can customize this as well if you want this to run at certain intervals now beyond that i didn't mean to back out but also adjust the time when exactly you want this to happen any notifications do you only want to get notified for important updates or just
failed runs only so it's completely up to you now again i would just recommend to be using this chat section just talking in plain english describing what you want you could also use this little microphone right here to dictate what you're saying to make it a little bit easier obviously because we speak faster than we can type at least 99 of the world now one thing that i do want to discuss a little bit further that i almost glossed over is going to be where this actually runs on so whether that's going to be locally on this device or on the cloud now if one is going to be running on your computer it can see your files and use everything we just set up but it is only going to be running while the application is open and your machine is actually on the other one it just runs in the cloud so it goes even when your laptop your computer is off and the cloud one can also fire off when something happens not just a timer maybe when a new email comes in from a lead it drafts the reply when somebody posts inside of a slack channel it's going to log it or do whatever it needs necessary that part you set up on the web or on your phone
not in the desktop application which is a little annoying but it does work now one of my favorite features about this if we go into our settings and we just go to phone we can easily connect any other device to this so what i typically do is i connect my phone and my laptop to this because obviously i'm not always sitting at my desk sometimes i travel and i just want to work from a different location whether it's a coffee shop or even a different state and i want to be using it or if i'm at the gym using it in those settings where i can just use my exact chat gpt setup look for all of my chats and recent things that i've been working on and operate from there simply just have to add your device so it's just going to pull up a qr code you scan that on your phone or on your computer you just enter in a code it is incredibly simple and all this allows you to actually work remotely one last thing that i do wants to be covering and it's more about where all of this is actually going so at this point you saw recorder skill watch the screen the bigger version of that is called computer use where astro just takes over the screen and it clicks through any of your applications for you but really any of the tools that do not have a plug so
if your crm does not have an api key that's where you can use a headless browser of course there's not really going to be any apis or anything for filling in a form online or updating a record in some old system like your crm or downloading a report from a website in openai they actually say that it's about twice as fast as this as the last model it actually just runs in the background while you keep working so that's where you could just use the computer use it'll control your computer do everything for you you could also just log into a browser right here it'll control everything for you so goes to google And I want it to search everything up for me.
I can log into my CRM from here. It'll control all of that without needing an API. There is some pieces where this does fall short and it's just as important as everything else.
But that's just to say that this is not a magic button that runs your company while you're sleeping. It most certainly can at this point. Every one of those 98 % numbers from OpenAI is self -reported and even they describe it as people running well -defined jobs through it, not just handing over the company, handing over the reins.
And I'd say the same from my experience. is fantastic at specific jobs, well -defined jobs for that matter. And it's not ready to be left alone as some general manager, at least as of right now, which is exactly how I'm telling you to use it.
Have real jobs, defined functions, and you reviewing all of the important ones. Now, two quick things beyond that. Number one is going to be the cost.
Now, work tasks, they're metered by how big the job actually is. And the workers, they use a lot more under the hood than just a quick chat does. In OpenAI's own help page, they say that heavy users landed about $100 to $200 a month per person.
And that's not it being broken. That's just what running a team actually costs. So start small, watch all the usage, and scale what is clearly working.
Now, the second thing is garbage in. garbage out if your files are a mess and your notes are chaos your team's output is going to be messy too and the businesses that are going to be winning with this feed it clean organized information so do a little tidy up it will pay off exponentially for you and the golden rule on all of this is just picking one workflow get it working three times in a row and then you can expand from there into adjacent processes or departments and the people who are just trying to automate everything at once they give up their team doesn't use it and they go back to doing it by hand and complaining that ai doesn't work it's a scam and they're never going to use it again and we see it over and over again the people who nail one thing are the ones who actually end up running their whole operation on this.
Now, all that to say, to kind of just wrap things up, the real change with this isn't a faster tool, it's that your job is going to be changing. You are going to stop being the person grinding through the tasks and you become the person who is directing the work and checking the outputs and you move from doing to deciding.
And that's the whole point of running a team. And you just built this one out of a simple project in OpenAI. You have a few skills and a recorder instead of hiring for every single seat and ramping up your costs in.
you know, onboarding. Now, if you're a business owner and you'd rather not build this and train yourself on this, we have done this for just under 100 companies at this point. And the goal every time is the same.
The business goes AI native, which basically means the processes are actually running on AI and your team directs the work instead of grinding through it. So if you want to automate your processes and transform the business the same way, there's going to be a link down below in the description to learn more about working with us and seeing if you qualify.
Now, beyond that, if you're wondering where to start. Every single business that I have worked inside has the same three leagues, chasing people for things. So whether that's documents or approvals or any replies, moving information between your systems by hand.
and reports that somebody rebuilds every week that nobody reads until they are wrong. And that is where the hours go, right? And the easiest way to actually find which of those is the biggest one for you is the free AI business growth plan.
And the reason that this exists is because most owners actually automate the task that annoys them the most, not the one that is quietly costing them the most. So you just answer a few quick questions and it tells you the systems to build, where you are right now, what it looks like after in an honest range on the hours and dollars that you would be getting back.
It just takes a minute to fill out. It's completely free and it's yours to keep either way. You can get that down below in the description.
The Hook
The bait, then the rug-pull.
Nick Puru opens with a blunt reframe: most people still think ChatGPT is a browser tab you type into and copy answers out of, and that single wrong assumption is quietly costing businesses hundreds of thousands of dollars. What follows is his full setup for turning GPT-6 Astra into a team of AI workers with memory, saved skills, and real tool access instead of just a chat window.
Frameworks
Named ideas worth stealing.
03:07list
Three Problems With a Normal AI Chat
It forgets you the second you close the tab
It cannot touch any of your real tools
It only works while you are sitting there typing at it
The three structural limits of a plain chat window that the rest of the video's setup is built to fix, one piece at a time.
13:35list
Record a Skill
Do it once, on your screen, the normal way you always would
It watches
It writes the skill
You say the word and it runs
The four-step loop for capturing a reusable skill from a screen recording instead of writing instructions by hand.
23:20list
How You Stay In Control
Plan mode
Approvals
Permissions
The three levers that keep a human in charge of an AI system with access to real tools and files.
22:00model
Your Departments
Proposal worker
Content worker
Operations worker
Research worker
A set of narrow specialist workers, each confined to its own project so it only carries the context its job needs.
32:00list
The 3 Leaks In Every Business
Chasing people for things (documents, approvals, replies)
Moving information between systems by hand
Reports rebuilt every week that nobody reads until they're wrong
The three recurring places business hours actually go, offered as the starting checklist for deciding what to automate first.
CTA Breakdown
How they asked for the click.
VERBAL ASK
32:10link
“The easiest way to actually find which of those is the biggest one for you is the free AI business growth plan... it's yours to keep either way. You can get that down below in the description.”
Soft-pitched twice (once mid-video around the pricing section, once at the close) as a free diagnostic quiz rather than a hard sales pitch, framed as identifying the viewer's biggest time leak, which lowers friction to click.
Add Modern Creator as a preferred source and Google shows you more of our breakdowns in Search, Top Stories, and AI Overviews. It only changes what you see, and you can undo it in your Google settings anytime.
Add to Preferred SourcesOpens your Google source preferences with us pre-loaded. Tick the box and you're done.
A live walkthrough of the manager-plus-specialist Claude agent system one founder built to run his content brand, his 20,000-member community, and his automation agency from one folder.
A four-step model for building free AI websites with GoDaddy Airo, turning them into $300/month retainers, then upselling automation and AI consulting to local businesses.
A 14-minute walkthrough of the Noose desktop installer that finally lets non-technical users run one of the most capable open-source AI agents without touching a terminal.
A 17-minute walkthrough of Max Hermes: the cloud-hosted Hermes agent that costs 95% less than Opus 4.7 and writes its own skill playbooks after every task.
00:38
01:11
02:23
02:52
03:20
03:43
04:01
04:41
05:17
06:10
06:57
07:43
08:36
09:37
10:25
11:13
11:57
12:27
12:49
13:49
14:32
15:28
17:47
18:53
19:42
19:53
20:13
20:39
21:02
21:52
22:23
22:53
23:27
24:03
24:39
25:15
26:17
26:46
27:22
27:51
28:05
28:36
28:55
29:30
29:44
30:13
30:23
30:32
30:51
31:14
31:30
31:44
32:18
32:30
19:18
37:52
18:51
24:57
14:31
17:04