A live, warts-and-all build of a Facebook-ad tool in Codex, used to walk through seven stages from one lazy prompt to a self-improving AI agent.
Posted
1 weeks ago
Duration
Format
Tutorial
educational
Views
13.2K
210 likes
57 · 43
Big Idea
The argument in one line.
Moving from basic AI prompting to a fully autonomous system takes seven deliberate stages: planning, a persistent file-based memory, live data connectors, reusable skills, and finally a cloud-deployed, self-improving agent.
Who This Is For
Read if. Skip if.
READ IF YOU ARE…
You use ChatGPT or Codex for one-off prompts and want a repeatable system instead of starting from scratch every session.
You run a business (ecommerce, agency, coaching) and want AI to produce on-brand creative or content using your own data, not generic output.
You're comfortable setting up folders, APIs, and basic automations, or willing to ask the AI itself to walk you through it.
SKIP IF…
You're looking for a no-code, point-and-click AI tool; this walkthrough lives inside Codex's file system and API connectors.
You want the finished Facebook-ad tool handed to you; the creator openly shows the build staying rough through the whole video.
TL;DR
The full version, fast.
The video lays out seven stages for leveling up AI use, from typing a single vague prompt (stage one) to running a self-improving, cloud-deployed AI agent (stage seven). The core mechanism is building a persistent file system around the AI, an agents.md ruleset, a memory file, a change log, and backups, so each session accumulates context instead of starting cold, then feeding it live data through API connectors like Apify so its outputs reflect real competitor ads and real business information. From there, repeatable tasks get turned into named skills, skills get chained into agents that run on a schedule or in the cloud, and the end goal is a loop where ad performance data feeds back in to improve the system on its own.
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.
Diagnoses the one-vague-prompt failure mode and previews the seven-stage system, built live around an AI Facebook-ad tool.
00:42 – 01:09
02 · Stage 1: the baby stage
A single context-free prompt into Codex produces generic ad-angle ideas that don't account for what's already running.
01:09 – 03:24
03 · Stage 2: the planner stage
Uses Codex's plan mode and a plain-language goal to get a structured six-step plan before any execution starts.
03:24 – 09:05
04 · Stage 3: building the AI's file system
Sets up agents.md, memory, logs, and backups so the AI carries context between sessions instead of starting cold each time.
09:05 – 15:11
05 · Stage 4: data hygiene and API connectors
Connects Apify to scrape a public Facebook ad library for competitor inspiration instead of logging into a live ad account.
15:11 – 21:19
06 · Stage 5: becoming a skill builder
Turns the repeatable ad-creation process into a named skill, iterating through rough first drafts toward a usable batch of creatives.
21:19 – 25:29
07 · Stage 6: autopilot
Compares a local cron job against deploying the whole tool to Netlify as a shared Kanban-board SaaS app, plus GitHub for team collaboration.
25:29 – 27:01
08 · Stage 7: loops and self-improvement
Closes with feeding ad performance data back into the agent to self-correct, and a caution against giving AI full unsupervised autonomy.
Atomic Insights
Lines worth screenshotting.
Most AI use plateaus at stage one: one vague prompt, one generic answer, then a verdict that AI doesn't work.
A browser chat session has almost no persistent memory; a desktop AI tool with its own file system can hold context across an entire project and business.
An agents.md file acts like a standing rulebook the AI rereads on every task, so explicit bans stick instead of needing to be repeated every session.
A dedicated memory file, a change log, and automatic backups turn 'why did the AI suddenly change everything' into a problem you can actually debug.
Giving an AI tool read access to a whole Documents or Desktop folder means it can see, and potentially affect, everything in that folder, not just the current project.
Using a no-login API marketplace to scrape a public ad library is safer than connecting an AI agent directly to a live ad account, because AI-driven API loops have gotten real ad accounts permanently banned.
Asking an AI tool to generate three to five variations of an output and picking the best one beats asking for one perfect result on the first try.
A reusable skill is a named, repeatable process that replaces re-teaching the AI the same steps every session.
Every first attempt at an AI-built tool or ad batch is expected to look rough; the point of a skill is that each flaw gets patched into the process so it isn't repeated.
Stacking three skills, data analysis, then creation, then a dedicated checker, catches mistakes a single skill running alone would miss.
A cron job only runs automation while the laptop is open; deploying the same workflow to the cloud keeps it running and lets a team use it like shared software.
The final stage closes the loop: ad performance data gets fed back to the agent so it can propose changes to its own skills instead of a human re-prompting from scratch every time.
Giving an AI agent full unsupervised control of a computer is something the creator actively advises against for almost everyone.
Takeaway
Seven stages separate a wasted AI prompt from a self-improving agent.
THE 7 STAGES
The jump from typing one vague prompt into ChatGPT to running a cloud-deployed, self-improving AI agent comes down to seven deliberate upgrades: planning, persistent memory, live data connectors, reusable skills, automation, and a feedback loop.
02Stage 1: the baby stage
Even a lazy, context-free prompt isn't useless, it just produces generic ideas that don't account for what you're already running or what's already failed.
The gap at stage one isn't the AI's ability, it's the missing context: current campaigns, past changes, upcoming promotions.
03Stage 2: the planner stage
Before building anything, ask the AI to plan first; a 'plan mode' or the single word 'plan' stops it from diving straight into execution.
State the goal in plain language, including your own skill level, so the plan it returns matches what you can actually follow.
A good plan surfaces the real sub-problems up front, like data upload, image generation, and which connectors to use, before a single line of work starts.
04Stage 3: building the AI's file system
A folder-based file system beats browser chat because it gives the AI real memory and the ability to pull from dozens of documents instead of one disposable conversation.
An agents.md file is a standing rulebook the AI rereads on every task, so hard constraints don't have to be repeated every session.
A separate memory file and a dated log turn 'why did this suddenly break' into something you can trace back to the exact prompt that caused it.
Automatic backups mean an AI that accidentally overwrites your data has somewhere to restore from, and advanced models will check the backup folder unprompted.
05Stage 4: data hygiene and API connectors
Data hygiene means designing how fresh, real data gets back into the AI regularly, not just setting up folders once and walking away.
Think like hiring for the role: what does a competent employee in this job download and study every week to get sharper outputs?
Use a no-login API marketplace to pull public data like a competitor's ad library, instead of connecting an agent directly to a live ad account.
AI agents making unsupervised, high-volume API calls into live ad accounts have gotten real accounts permanently banned; a read-only scraper avoids that risk.
06Stage 5: becoming a skill builder
A skill is a named, repeatable process you define once, instead of re-explaining the same steps every session.
Expect the first output from any new skill to be rough; fold every specific flaw you spot back into the skill's instructions so it's fixed next time, not every time.
When generating creative output, ask for three to five variations side by side rather than one 'final' version, then pick and refine from there.
Stacking a data-analyst skill, then a creation skill, then a dedicated checker skill catches mistakes that a single skill running alone would miss.
07Stage 6: autopilot
A cron job is the simplest automation, a scheduled task that re-runs your agent through its skills, but it only works while your computer is on.
Deploying the same workflow to the cloud removes the 'laptop has to stay open' dependency and lets teammates use it like shared software.
Putting the project on GitHub once you're deploying to the cloud lets a team see the codebase and suggest changes instead of only one person holding the context.
Cloud deployment comes with real security tradeoffs; ask the AI directly at the end of a session whether it spotted any security risks that day.
08Stage 7: loops and self-improvement
The final stage connects outcomes back to inputs: real performance data gets fed back to the agent so it can propose changes to its own skills and agents.
Treat agent-proposed changes as suggestions you approve, not changes it makes unilaterally, especially while you're still trusting the system.
Full-autonomy agent tools with open access to a computer exist, but the creator explicitly says he would not recommend that setup for 99.9% of people.
Glossary
Terms worth knowing.
MCP (Model Context Protocol)
A standard that lets an AI agent connect to external tools and data sources, such as ad platforms or scrapers, as callable connectors rather than one-off manual integrations.
Apify
A marketplace of prebuilt data-scraping tools that an AI agent can call through an API, used here to pull public Facebook ad library data without logging into any ad account.
AI agent
An AI session set up with its own context, rules, and a set of skills so it can carry out a multi-step task on its own, rather than only answering single questions.
cold open thesis line, sets up the whole video→ TikTok hook↗ Tweet quote
09:51
“What is the process that I do each week to improve my intelligence? What is the data that I download as an employee to change the outputs that I'm creating?”
reframes an AI workflow as 'what would a good employee do', concrete and quotable→ newsletter pull-quote↗ Tweet quote
16:51
“Pretty much every single time, the first attempt is going to look terrible. This is the whole point about building these AI skills.”
manages expectations with a counterintuitive reassurance→ IG reel cold open↗ Tweet quote
19:28
“My top tip with AI: if you're trying to generate things, get it to create three to five of them, and probably one of them is going to be good.”
tight, actionable tip with no setup needed→ TikTok hook↗ Tweet quote
26:13
“For 99.9% of people, I would not suggest doing this.”
unexpected caution from a heavy AI user, a pattern interrupt→ IG reel cold open↗ 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
analogy
Most people are actually using AI wrong. They'll open a chat, type one vague sentence like, give me Facebook ad ideas. They'll get a few generic concepts and then decide AI is just not useful.
In this video, I'm going to show you how to go from stage zero to stage seven, which is an absolute master using Codex and ChatGPT. This is the exact leveling up process that I've done firsthand. all of my AI tools and workflows, which have now helped me do over $1 billion in sales across my businesses.
So today I'm going to walk you through how to build the blueprint and actually create an AI image ad system for your brand. By the end, I want you to feel like an elite expert that knows instantly how to use AI to solve any other problem in your business. By building a simple tool like this, you're going to learn all of the foundations.
So stage one is the baby stage, which a lot of people are still doing. They simply go to a tool like Codex and they simply prompt at something like, I sell a wearable blanket, like the Udi. I want you to create 10 Facebook ad ideas to help me improve my Facebook ad results.
Then they submit it. You can see here, it's given a bunch of ideas, such as I refuse to take this off angle, the heating bill angle, texture close -up ads. Now, these actually aren't bad ideas.
However, they don't actually help us get the Facebook ads results quickly. They also aren't considering a ton of other contexts, such as maybe I'm already running these ads. Maybe I've changed my product.
Maybe there's a particular promotion that I should be running sooner and focusing on. There's just so much context that Codex doesn't currently have that would make our results so much better. So to improve this, we move on to stage two, which is actually the planner stage.
Every year I take hundreds of mentoring calls and I'm always shocked on how many people get stuck on a very, very simple process simply because they fail to either plan with AI or just ask AI how to get unstuck. So to do this, what you want to do is come to these projects. Let's create a new chat here.
Now I can turn on this plan mode if I want to make sure that it doesn't do anything and it doesn't really just dive in the deep end. Or I can actually just prompt it to actually start planning by using the word plan only. Now, it's really important whenever you use AI to really understand what you're trying to solve.
If you can't tell Codex what you're trying to solve for in the very, very early stages, you probably need to go back to the drawing board and think about it. But for me, I know exactly what I'm trying to solve. So I'm going to prompt this now.
I'm a beginner. Help me plan out this project that I want to do. What I want to do is I want to create a Facebook ad tool that understands all of the context of my business, creates winning Facebook ad ideas, as well as even creates the static image ads for my Facebook advertising account.
This is so I can create profitable ads. Now, if we really wanted to, we could simply say, I'm not technical. Keep the language very, very simple.
And then I can submit that. Now you can see here it's put together a pretty simple plan that I think anyone could follow. It understood that you need to build a tool where you upload things.
It understands that it creates an image. It understands things like the ad library as well where all of the ads are saved. It even talks about certain connections which we're going to talk about in a second such as APIs and MCPs.
And it also even talks about finding real winners. So realistically, you can now see this exact six -step process that it's talking about. And then you can start to challenge it.
And by the end of that, you're really going to have a good tool. Now, if I really wanted to, I could probably just tell Codex now to go ahead and implement this and walk me through each step. You'd probably end up with a really good tool.
But remember, I want to teach you the foundation so you can build any tool. So let me talk about the next fundamentals. Now, if you actually are following step by step for me because you actually want to create this tool that I'm talking about here, be aware you probably are getting different results on your screen right now.
But because you might be using a different model, it might be have a little bit of different context, all of that kind of stuff. So don't stress. Just keep learning these fundamentals with me.
If you get stuck and something's not making sense compared to my tutorial, just download the transcript of this video and put it into your codec session and say, I'm getting stuck at this section. When I was playing around with AI maybe six months ago, I was really struggling doing complicated financial models. I was trying everything like using the Gemini extension with Google Sheet.
I was trying to use things like ChatGPD, deep research, thinking mode, and just nothing would work. It wasn't really until I started using Codex and actually using the file system on my computer that I had enough power and enough understanding of all of my projects. Then it finally worked and I could build out incredible things.
So the way that a lot of people are using AI is they'll go onto their Google Chrome session and they will start chatting in the browser such as onchatgpt .com. The problem with this is it has very little memory and it has very little context. It also has far less juice the ability to actually process information.
Whereas when we use desktop apps like Codex and Chord Code, now we can use a file system that we create rules around as well as give access to all of our AI apps so it can consistently keep it up to date and get as much context as it needs. This includes things like memory, save logs, even backup files. But the main thing is it's able to pull on 50 documents at a given time, really understand the context of a whole project and the whole business even, and have far better answers.
So to get the files set up correctly, what we're going to do is we're going to go into this chat and we're going to say, can you please make the proper file system to get the best use out of Codex, such as the agents .md file. I'll explain this in a sec. Memory.
logs and backups. Then I'm just going to paste this and while it's setting up all of those folders, I'm going to explain what they each are. Then I'm actually going to show you where they are on the computer.
Now, I know this is feeling a little bit technical, but trust me, this is super, super powerful and you'll understand it as soon as you start using it. The agents .md file is similar to a chord .md file if you're using Chord. In short, every time that Codex actually does something, it's going to read the agents .md file.
So you can give it very, very strict, explicit instructions to avoid things and making it go crazy or doing the wrong thing. from hallucinations. It might be for us, do not make false medical claims on Facebook ads.
It may also give it instructions about where to find connections to other tools, where to find really important data, or even how it reads the data across our folders. Think of it as a prompt that always kind of goes into every single new chat. Next we've got memory.
AI is almost always using some sort of memory. However, by creating the folder structure and creating memory as a clear MD file, which I just currently did, it allows you to actually read the MD file and go back and say, oh, no wonder it's making these mistakes because its memory thinks that that's what it should actually be doing.
Then we've got logs, which is very similar to what I just said. But in short, you could see, oh, I prompted it to actually do this a long time ago. And that's when everything started to go wrong.
Maybe last week I told it to actually change the way that it renders copy on the static image ad. completely forgot. And then I'm suddenly wondering, why is everything starting to change?
I can go through my logs and say, this is when things started to go wrong. Please change it. And then finally, we've got backups, which I highly recommend using.
You can use something like your agents .md file to say, always write to backups regularly. So let's say that you're creating a spreadsheet with all of your top hooks from Facebook ads and then suddenly your AI just completely overrides them all. You could then go to your backups and you'll have that file in there.
In fact, the AI will probably go if it's missing context, it will check in the backups and will store from there without you even having to prompt it. So it's worth showing you exactly where this is showing up on the computer. Now, if you didn't select a folder at the start of Codex, it's just going to be in the default folder and it's probably going to say ChatGPG or Codex.
Now, if you're a beginner, this isn't necessarily a bad thing because then the AI is actually going to be restrained. Depending on what level of access you've given it, it can only probably see folders within that folder. So I'll just open this up in Finder.
So you can see here that it's created the folder just in the ChatGPT section. If I click on Documents, I've got this ChatGPT folder here and I can go into this project. And this new project is exactly that.
So when we open this up, you can see all of these files have now been created. It's even created a couple of extras so we can see things like decisions, project memory. I can open that up and it's got all of my goals.
Now you don't actually need to touch any of this stuff anymore. Let's say you did a big session one day and you're talking to Codex all day. You can simply say at the end of the day, you could go save to memory and logs and then make sure the agent MD files understand what we need to work on next.
It's going to change all of those folders for you. But truthfully, Codex is pretty good at updating them over time regardless. As I said, the cool thing with AI, if you're confused where this folder even is or how to use folder systems, just ask the AI itself.
Now, I should also say here that if you're starting to get quite advanced with AI and you want to give it more access to your computer, you can actually give it a higher level of access in these projects by selecting this source folder here. And you could do something like your documents. And it's going to be able to see pretty much everything in there.
I could also even select things like desktop. Eventually over time, you're going to constantly get asked prompts where it's asked for levels of access across these files such as downloads. And if you accept that, it's going to be able to see and access all of those files.
So you just need to be comfortable with that. But truthfully, the recent models are really smart around what it should access and what it should affect. Okay, so moving on to the next step, which is data hygiene.
I know this sounds technical, but we're actually going to start building out this tool in a very, very cool way. Now, data hygiene is one of the most important things in AI. And what I mean by that isn't just these folder structures that I've showed you.
What I mean by that is how do we dynamically change what the AI is actually looking at so that we can create a sustainable, repeatable process that doesn't give us the same bland results every single time. Now, what I recommend if you're not building this tool alongside me and you're building something else, what you need to understand is if I was an employee coming in and focusing on this role, what is the process that I do each week to improve my intelligence?
What is the data that I download as an employee, as a person to change the outputs that I'm creating? For Facebook ads, let's say a creative strategist, what they're going to do is they're going to go into the Facebook ads library. study competitors or even look at your top performing ads across other tools.
And then they're going to use that fresh data, they're going to digest it and then they're going to adjust what they're actually outputting. So we just want our AI to do that. And this is the perfect time to start talking about APIs, MCP, CLI and all of those other type of connectors.
Now, as I said, if you're not building this tool with me, you can simply prompt in, how do I make sure that my data that comes in every single week or every single month updates dynamically? And then what type of APIs or connectors should I actually use? However, for me, I know exactly what I need to do here.
All I'm going to do is, can we use API Fire? Now, if you're a bit confused about what that is, I'm going to explain it in a second. It has a Facebook ads library scraper.
I want to use that to actually digest and analyze ads that will be the top performers for me. I'm actually going to build this tool for Daily Mentor. That's another one of my businesses.
Let's start by analyzing Daily Mentor Facebook ad account library via the ads with the most impressions. Just analyze the top 10 creatives for now. Now, what I mean by this is there's a tool called API Fire or maybe it's called Appify.
And you can come on here and see that there's all of these APIs that I can now connect into. Now, could I build a scraper that goes into Facebook Ads libraries without me logging on? Maybe.
In fact, definitely. Could I build a public scraper that understands Reddit, that uses proxies, does all of this complicated stuff in a daily business and build my own tools on here? Of course I could.
However, I would much rather just tell it to connect to someone that's pretty much built all of these things out, which is this API file that I'm talking about. and then have the capability straight away. It's very, very simple and it's also very, very cheap.
I'm in no way sponsored by these guys. I just always default to them when I'm testing a new AI tool because I can just connect into their platform. So all we need to do here is you go to the console, just sign up.
Now to load most of their tools, all you need to do here is go into settings and copy this personal API key. Now, as always, if you're confused about how to connect a tool, simply go into Codex and say, I want to connect this tool. How do I do it?
It will even give you where to find it on their website. So you can see here, I've got this personal API token. I can simply click copy here.
Now you don't want to paste this into Codex. If you really want to be strict, you can simply come here and say it's copied to clipboard. And the reason why we don't want to paste any keys in because these are actually really, really secret and we don't want them leaked.
But if you're using Mac, what this is going to do, it's going to save it into the keychain and it's going to be able to access it later in a secure way. Now, you might be wondering, why am I not using the direct connection into Facebook for this research process? Truthfully, I probably can now.
However, just recently, there was a wave of bans for Facebook ad accounts because people don't really know how to use AI properly and they were using way too many API calls into the actual connection into Facebook. This got their ad account permanently banned.
It was probably just AI constantly looping and they were launching thousands of creatives every single day and Facebook didn't really like that. So I'd prefer you as a beginner to just simply use accounts or APIs that doesn't require you to log in because I don't want you to get banned. So you can see here that's now run and created screenshots of all of my top performing ads.
I think it's grabbed a couple of videos, which is probably not what I want. I probably just need to refine it and go back and say, only look at static image ads. But you can see it's actually starting to understand my Facebook ads and can analyze it a little bit.
Now you can see here we've got the scraper working. I've added Alex Homozy and Sabree, my fellow Shark on Shark Tank. These guys are definitely the elites of the elites.
They have great teams, their static ads are amazing. We're obviously not going to be copying them directly. Our AI is just going to be using them as inspiration.
And we may use similar things like the text format or something similar like their angles. They're also not direct competitors in my space, so shouldn't be an issue. Truthfully, looking at the tool, it hasn't pulled enough ads.
It's duplicating the same ad across all of my statics. And Sabree, you can see here he's got this news one, which is a great format that the AI will actually pick up on. But a lot of them here are actually just screen grabs, which is actually an insight in itself.
But I'd want it to go a little bit deeper and maybe grab another 20 images. Or I would go and add another 10 Facebook ads libraries of people with great inspiration. So I'm going to keep going with the AI back and forth pushing through with this tutorial.
And I wouldn't be surprised if it keeps getting a little bit confused, but every time it does something weird, I'm just going to go back to it and give it a prompt. Now, what I've actually done now is connect something called a product Bible. This is basically just a massive voice note that I recorded about everything what my product is.
I've then put that in my folder system as well as some information about things like our sales process as well as the objections that we often get around the business. And this should help refine it and allow it to create our first proper static image ads. Now, one thing I can actually do here is come to ChatGPT or Codex and simply prompt in start creating ads for this.
Now, after some back and forth, I feel quite confident I could actually start creating some ads. However, what I would highly recommend is learning the next stage, which is stage number five, become a skill builder.
Now in AI, we have agents. This is your AI agent that does things, that understands all of the context, that writes your memory logs that we talked about before. You're going to be able to automate that agent.
You're going to be able to get that agent to talk to other agents on your computer. However, what we really want to understand as well is skills. Because skills are very easy.
You can create a repeatable process where I go, use the Facebook ads static tool skill and create me 10 new image statics. Then it will go through a repeatable process that you've programmed so you don't need to constantly prompt it and teach it new things. Because if you keep doing that, eventually it's just going to get confused and start drifting.
Then you start getting frustrated and then you say, AI doesn't work. So to solve this, we're going to use skills. So to create the skill, I'm just simply going to say, let's create a skill, the Facebook ad static tool.
I want you to analyze the sales data. analyze my product Bible. Then I want you to scrape at least 30 Facebook static ads from the inspiration people.
I want you to find some patterns and create some cool ideas. Then I want you to sign off on those ideas. Then I want you to go and continue with a skill and create some static image ads now.
Honestly, now I think about it, I could probably have just said, create a skill with all of the things that we've just done before so I can create Facebook ads every single week. However, I can just submit this and it's going to create the skill. So now what I could actually do is I could come into Codex and I could say, use the Facebook static tool or I can simply forward slash and actually write the actual skill in here and then it should start running the exact process.
Now believe it or not, I'm actually not too hopeful about our first ads. The key with AI is pretty much every single time, the first attempt is going to look terrible. Especially if the tool is somewhat complicated.
Even when you deploy something, a brand new website, there are always going to be buttons or things out of place, things that don't click through or there's just going to be some bad copy or the logo is going to be in a weird position. There's just always something a little bit off about it. This is the whole point about building these AI skills.
They're not instant, but as we build on them, they're repeatable and you can simply automate them as well. Every mistake you see, you can build it into the skill by prompting in and next time it won't make it. So you can see here, it's created this reusable skill.
I should be able to go Facebook static tool, run this skill and it's going to run up. Now another reason I don't think this is going to actually work too well is I actually want to link up an MCP into Higgs field as well, which is a much better image generation tool. ChatGBT is actually quite good at rendering text.
even logos on your ads at the moment, but I found having something like Higgs field is really, really great for image generation. The other thing that I haven't really given it yet, which I actually probably will do now is the assets that it actually needs to create the ads, such as a photo library of maybe me or my Instagram.
There's probably going to be a couple on the internet. So what I'm going to do is I'm actually just going to get it to pull of all of those photos for me. But you need to understand where all of your images are.
Maybe they're all on your Instagram. Maybe you've got some folders. Maybe you've got Dropbox or a Google Drive somewhere that you can connect into your Facebook ad tool.
Higgs field, if you want, can really make robust AI images, understanding your product dimensions. You could give it a few 3D renders. You can give it a ton of context and actually start building the ability to create your product within Higgs field.
I've got some really good video on that free on YouTube if you want to learn it. However, at the moment, it's going to be much easier when we're building a tool like this to just have some really good photo images initially so that we know that that's not the variable that's stopping us from creating profitable Facebook ads.
And now you can see here that it's given our first few ads which look actually very similar to our already top ads. It's probably just understood the ad and just gone way too hard. thinking about that one.
You can see that it's grabbed some of the structure there, but you can see it's grabbed the logo, which makes no sense. It's got the background. You can see that it's got things a little bit cut wrong.
You can still see that it's grabbed screen grabs of my videos. I actually forgot to prompt and fix that. And you can also see that they're just all the same ads.
So this is exactly why we do skills, as I said before, because now all I'm going to do is go back to it and say, you need to make sure that you grab the actual daily mentor logo with transparent background. It should be in the folder structure. can go when presenting your ads, I want you to make sure that you present the example inspiration as well as show three of our ads next to it so that I have a few things to choose from because this is my top tip with AI.
If you're trying to generate things, get it to create three to five of them and probably one of them is going to be good. Then I can say, make sure you provide a batch of ads with multiple different inspirations because you've delivered ads that all look pretty much exactly the same. Make sure you change the skill to match all of this feedback and make sure all of those things are checked before delivering the next batch and then I can submit that.
And we can give that a little bit to load. Now we can see that it's actually fixed these things. It's obviously got a JPEG for the logo so I can fix that by removing and giving it a proper logo.
I probably wasn't clear about that or I didn't even have it in the folder structures. That's definitely my fault. And it's created option A, B and C for the inspiration.
Now these ads are definitely not perfect and I could probably get a... perfect maybe with another one hour or two hours of finding the skill, but I just want to finish explaining skills and agents before moving into the final two sections. So we have obviously the creation tool, the skill there, and part of this skill is obviously checking the logo, that kind of stuff.
However, if I really wanted to get regimented, this is what I could do, is that I could create a second skill. after this first skill that is a checker and I could literally come to this codec session and say run the checker on the most recent batch of ads that you just created. It might be a batch of a hundred creatives and it might find a lot more problems that was not originally caught in the first skill.
I might even have a skill before it which is called a data analyst skill. And it goes through and it tries to qualify all of my results that I previously got across ads libraries, other people's ads libraries.
It looks at data. It really understands the data component of the actual process. However, what I'm trying to say is you can have multiple skills.
Then on top of it, I can simply have an agent. So I can go to this agent and say, go do these three skills in order every single Monday. And then it's got all of the context.
It's got all of the processes. It can hand off through those skills and go through a regimented process, which then brings me to my next point, which this allows us to really focus on automation. So the next stage is autopilot.
So to discuss automations, let's just recap quickly. So we've got the agent and we've got three skills underneath that is a handoff process. As I said before, you can definitely go to the codec session and say, I want you to automate this agent through a cron job.
A cron job is just a scheduled task that it's going to do. Your laptop needs to be open. The agent's going to pull on the skills as you've kind of told it to, and it's going to produce the result.
Truthfully, that's definitely the most basic form of automation. It's the easiest one to set up. However, it's not the best in my opinion because I manage a lot of teams, a lot of people.
I want other people to be able to use my skills. I want them to be able to see my skills. I want them to be able to use all of the tools and processes that I set up.
What I actually do is I deploy it to the cloud. So I'll use something like Netlify. You can use whatever your codex suggests.
You see, all of the stuff that we just kind of did is running on your computer and it's all just code. What we can do is we can deploy it to the cloud so that people can log in like their own SaaS tool and actually start using the tool accordingly. But the main reason that I really like doing this is if you set up cron jobs, you need to have codecs open for that automation to actually run.
If you're just using your own laptop, you might not always have your own laptop on, which means like things that are collecting data or trying to run every single day to improve our data hygiene might just not be getting updated. They might be stuck and you might not even see that. So I'd much rather say something like this.
I want you to deploy this on Netlify so that I can eventually give my team access to it. I want you to think really hard and creatively about how to present this as its own SaaS tool with really good user experience. My suggestion is a Kanban board where I can simply click into the actual item, see the inspiration post, see our three creatives that we're creating, give me five creatives in the Kanban board so I can see how it's actually going to work.
Also make sure that I can actually suggest changes and edit the images. Now one thing to note here is if you are deploying into the cloud, the security requirements do come a little bit different.
Now, if you're using a model like 5 .5, you're probably watching this a bit later and the model is even better. The security ability and its understanding will just be way better than it even is today and will probably be able to warn you about things like security. But you can always just prompt it and remind it, is there any security risks that you've spotted today at the end of your session?
Now, I just realized I did something a little bit stupid and kind of jumped the gun. That was supposed to be my next section because... I don't actually need to first deploy this Kanban board to Netlify straight away, especially for this tutorial.
So if you copied me with that, it's okay. You want to eventually get to that stage. But what I could have actually done is just deploy it as a local SaaS tool on my computer because then I could simply say, okay, deploy this on Netlify now, but I probably just gave you an extra signup step.
But as I said, eventually you want to get it on the cloud. So that's okay. So now's another good time to talk about one other thing that...
As you start to become better at using AI and start hiring more teams, you'll actually start using something called GitHub to deploy the code there as well. Once it's on GitHub, it will then deploy on Netlify so that you can actually see all of your code base and changes that you've previously done. You'll be able to invite all of your team to GitHub as well so that they can make commits or suggestions to the code.
And again, if you're confused about how to do this or what level of access to give your team and the workflows for that, go to your Codex session and ask it. Now you can see here, we've got a Kanban board where we can easily drag these creatives. You can see it's picked up on Sabree's awesome breaking news format that he's done.
I should give Sabree just a little bit of love here because we have publicized his top creatives. So make sure you go check out his social media. You can see there's a few formatting issues again.
And I could easily go down to this section and say, can you change this image here? And I'll probably do so for this. But honestly, it looks pretty good and it actually deployed in Netlify in the end.
So this brings us to our final section, which is loops and self -improvement. So in short, we've created Facebook static ads that we're hoping will make us profit. To create a proper loop and self -reinforcement, we need to connect the end result with our initial inputs.
So what that would look like is we're going to have to find a way to connect into the Facebook return on ad spend and allow that to come back and adjust the actual skills and agents that we set up in the first place. Let's say we launch all five of these ads and then none of them get any results. We then give it permission to suggest changes back to us to say, none of my ads got results.
I want to change this process, this process, this process. And we can say, good idea. go ahead.
Maybe one or two ads start to really perform or maybe it says that I just want to inject double the amount of creative so that I can actually get the results and then you simply set up a system where you say approve. Now, one system that you can use here is something like OpenCore or Hermes or Hermes, I don't know what it's actually called.
And this is actually sits on top of your computer and just has access to everything and it just goes wild. However, for 99 .9 % of people, I would not suggest doing this. Instead, what I would do is I would ask my codec session how can I actually create loops and self -reinforcement loops and self -relearning loops so that it can actually run the thing either on the Netlify deployed SaaS store or locally on your computer so that you and your AI can learn about things.
If you want to learn more about AI, check out the link in the description.
The Hook
The bait, then the rug-pull.
He opens by diagnosing the whole audience's failure mode: one vague prompt into a chat window, one generic answer, and a verdict that AI isn't useful. What follows is the fix, a seven-stage ladder from that lazy prompt to a cloud-deployed, self-improving AI agent, demonstrated live by building a Facebook-ad tool from scratch.
Frameworks
Named ideas worth stealing.
00:21list
The 7 Stages of AI Mastery
Stage 1: Baby stage (one vague prompt)
Stage 2: Planner stage
Stage 3: File system (agents.md, memory, logs, backups)
Stage 4: Data hygiene and API connectors
Stage 5: Skill builder
Stage 6: Autopilot (cron jobs and cloud deployment)
Stage 7: Loops and self-improvement
The spine of the whole video: each stage fixes the specific limitation of the one before it, starting from a single disposable prompt and ending at a self-improving, cloud-deployed agent.
Steal foran onboarding doc or training session for a team that needs to go from casual AI use to building real internal tools.
05:27model
The AI file system
agents.md: standing rules the AI rereads every session
memory: a running record the AI consults and updates
logs: a dated history of what changed and when
backups: a fallback copy the AI can restore from if it overwrites something
Four files that give a desktop AI tool persistent context across sessions, explained and set up live in Codex.
Steal forany Codex or Claude Code project that needs memory and auditability across long-running sessions.
19:40list
Skill, checker, data-analyst stack
Data-analyst skill: qualifies and digests incoming data
Creation skill: produces the actual output (ad creatives)
Checker skill: runs a QA pass on the batch before delivery
Chaining three narrow skills in sequence catches mistakes that one skill running alone would miss.
Steal forany repeatable AI content-production pipeline that needs a built-in quality gate.
CTA Breakdown
How they asked for the click.
VERBAL ASK
26:55link
“If you want to learn more about AI, check out the link in the description.”
A single soft CTA at the very end, no mid-roll pitch; the heavier mentorship-program promotion lives only in the YouTube description, not in the spoken script.
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.
Watch next
More from this channel + related breakdowns.
19:18
Alex Finn · Talking Head Tutorial With Screen Walkthroughs And Digital Whiteboard Diagrams
A screen-recorded walkthrough of turning a dedicated Mac mini and a ChatGPT subscription into an always-on AI operator that runs email, calendar, SEO, and daily priorities.
A two-hour, filmed-in-one-day walkthrough of everything Codex can do with GPT-6 Astra, from one-prompt iOS apps to an agent that reads your Slack and tells you what to work on.
A creator walks through the plugins, projects, and routines behind ChatGPT's agentic Work mode, then demonstrates it editing a video, drafting a sponsor script, and testing a video game all at the same time.
Riley Brown wires ChatGPT's new Work mode into eight always-on automations, from a daily commitments tracker to a fully scheduled monthly business review.
A walkthrough of ChatGPT Work's Sites feature: building a game, turning a project into an interactive site, wrapping a custom skill in an interface, and automating a self-updating newsletter database.
00:23
01:10
01:26
01:48
02:04
02:52
03:17
03:29
03:50
04:06
04:24
04:45
05:13
05:37
06:17
06:29
06:46
07:06
07:30
07:56
08:16
08:40
08:47
09:23
09:32
09:51
10:11
10:28
10:51
11:39
11:51
12:19
12:46
12:51
13:14
13:38
13:54
14:19
14:32
14:57
15:33
15:55
16:22
16:50
16:54
17:15
17:36
17:54
18:27
18:37
19:04
19:17
19:45
20:02
20:28
20:46
21:12
21:19
21:40
22:13
22:20
22:46
22:59
23:28
23:55
24:01
24:29
24:41
25:20
26:13
26:24
19:18
23:06
1:54:47
18:16
20:07
13:24