影片筆記：Learn 95% of Codex in 30 minutes

處理狀態

已完成語音辨識與逐字時間軸。自動摘要階段未在時間預算內完成，因此先發布可用版本；可於後續從現有逐字稿補做摘要，不需重新辨識影片。

來源

https://www.youtube.com/watch?v=474wZZHoWN4

完整逐字稿

You're in the right place if you use ChatGPT, Claude, Claude Code, Claude Co-work, or Codex, but don't know which one you should use for business and how.
Because in this video, we're talking about using AI agents effectively.
Specifically, we're talking about Codex, OpenAI's agent tool, which is like Claude Code, except it's an easy-to-use interface for doing knowledge work and coding work.
In this video, we're going to be breaking down the seven core capabilities of Codex, OpenAI's AI agent super app, each with real-world use cases, so that by the end, you know exactly [music] what it can do and where it fits in your workflow.
Let's get it fits in your workflow.
Let's get started.
So, this right here is Codex.
This is a clean interface for AI agents that can fully control your computer.
At first glance, it kind of looks like ChatGPT or Claude, because you can create a chat, but they're not exactly the same.
When you chat on ChatGPT or Claude and upload files, those files are actually stored in the cloud.
With Codex, all of the files that you upload to Codex or that you create with Codex are stored on your computer.
Codex is made to do any task that you might want to do on a computer.
You can create a motion graphics video.
You can create beautiful landing pages, and directly inside Codex, you can generate images that you can place in your site.
On Codex, you can also create more complex games and 3D simulations.
You can also create mobile apps with a front end and back end with a professional design.
You can also create desktop apps in just a few prompts.
You can ask GPT 5.5 to create you an Excel sheet after doing a ton of research, and it can create spreadsheets with multiple slides.
It can also create any type of Word doc with many different charts, and you can also create presentations directly inside Codex and export them to Canva or any other tool you want.
And I haven't even mentioned yet that Codex can fully control my computer, and it can also control the browser.
But, this is something we'll talk about a little bit later.
As we go through the main seven different capabilities of Codex, you'll realize that it is truly a super app that does any coding task or any knowledge work task.
Let's not waste any time.
Let's hop into capability number one.
So, the first capability that Codex has is full file access.
You see, when I use ChatGPT, all of the files that I give to ChatGPT or that ChatGPT generates are stored in the cloud.
However, when we use Codex, all of the files that are created or the files that I give to it are stored on your computer.
And the agent inside Codex has full access to your computer.
Let me show you.
In my downloads folder, I have a folder called receipts to process.
And we just have 60 different little photos of receipts that I've added to this folder.
I can simply ask Codex, "In the downloads folder, you'll see my receipts that I need to process.
I want you to analyze them in an Excel sheet and create charts that helps me visualize the types of transactions and any other patterns that you notice." So, we can choose our model, which is GPT 5.5 high, and we can run it.
And the agent, because it has full file access, it can access any file on my computer, it'll find the relevant folder and complete the task.
And you can see here, it found the target folder, which is users Riley Brown downloads receipts to process with only 53 receipts.
I'm going to OCR those, and so we're going to extract the text from those receipts, classify and categorize each transaction.
Now, it's creating an Excel workbook on my computer that we can open up in Codex.
All right.
So, after 7 minutes, it is now done and we can very easily just open this Excel project up directly inside Code X and you can see that it has created this file.
We can see there's a dashboard, the total spend is there's a dashboard, the total spend is 25,982.
We see receipts, category summary, payment summary, and we can see all the different trends, the spend by payment method.
We can see the monthly trends.
This is an Excel sheet and because it's on our computer, if you click this down arrow, you can see open in folder and we can see the exact file location of this Excel spreadsheet.
So, now that we know that all of the files created on Code X live on your computer, the question becomes how do you stay organized?
And I will tell you the best way that you can stay organized inside Code X is to work in a project.
So, you can create a chat in one of two ways.
One way you can create a chat is if you just go to this chats category and you just hit this new chat button.
It allows you to create a chat that just shows up here in the sidebar.
It's not actually inside a project.
Or, you can go ahead and you can create a project and you can always hit start at existing folder. all of my projects in my documents folder and then I will just create a new project seven capabilities cuz that's what we're talking about.
We can create a new folder and now we see that we're creating a new chat inside the seven capabilities folder.
You'll see here this project will show up here in the left side panel and as soon as I add a chat, you can see the chat shows up here.
The cool part about Code X is you can multitask and you can create multiple chats at the same time.
Here, I'm going to ask the agent to create a Word doc.
And we're doing this within a chat inside the seven capabilities folder.
Any document that the agent creates when it's in a chat inside a project will actually store the document that it creates in this folder on your computer.
So, if we click these three dots here and say open in finder and we open the seven categories folder, we'll see that there's nothing in it.
And that's because the agent is not done creating the document.
As soon as the agent's done, it's going to put as the agent's done, it's going to put the uh document that it creates in this project.
And as you can see here, we have this Word doc that was created and it's called learn 90% of OpenAI Codex in under 2 minutes.
You can see the title of the folder right here.
If we go ahead and open up finder at that location and we click on outputs and docs, you can see that that document is located right here.
So, this is literally on my computer by default.
And so, this video that I'm making now is called the seven capabilities of Codex.
And if we were to create a new chat in this folder, the chat would be aware that it is starting out in this project.
I can type at learn and see I can actually at mention any file that's created within this seven capabilities folder.
And I can say, "Please turn this into a landing page, a good learning resource for my viewers."
And so, I can take any document within this folder and I can just at mention it and say, "Please turn this to a landing page." And because it can control my computer and go look at all of the documents that are organized very
easily, it can find the right document and then create a landing page.
And the code for that landing page will exist on my computer.
And here we go, it's done.
We can open this up and we can see that it created a landing page based on that document that is stored on our computer.
And let's go ahead and move on to the next capability.
Next up, Codex has something called persistent memory.
So, within Codex, there's two types of memory.
There's manual memory, and there's auto memory.
Manual memory is when you ask the agent to remember something, when you tell it, "Hey, can you remember this going forward in the future?" Let me show you an example.
So, I made a few changes to that landing page and got it to create it in this format with this little left sidebar.
We can scroll down, and we can automatically snap to the section that we like.
I really like this format, and we like.
I really like this format, and so from now on, can you please use this format whenever I ask for a landing page and I don't specify uh the styling?
You should use this styling and remember that if there's a lot of text content to use this table of contents.
So, I can simply just ask the agent to remember this, and it's automatically going to add it to the agents.md file, which is where you can see here, search for agents.md.
This is where it's going to store this information for later.
And so, you can see here, the agent updated its manual memory, which is the agents.md file.
If we click on this, here are the things that I've asked it to add in the past.
Here's the what it just added.
It says for landing pages that are made for learning purposes, it says use this text first format, um and then it kind of describes the style that I like.
This agents.md file is a living document that gets updated over time that acts as the manual memory.
But, there's also a different type of memory that you should never and this is an auto memory feature that Codex kind of keeps updated automatically.
And just like the agents.md file, this is all stored in a file on your computer.
If we go to Codex and hit command N to create a new chat, please tell me about what you have in the and then Codex / memories folder.
I want to understand what you know.
Also, open that file.
And you can see here it gave me the path to this file.
We can go ahead and open this up.
And this is a document you should not touch.
You can just kind of observe it to see what the agent observes about you.
And so every few weeks I'll actually go through this and take a look.
And this actually gives me a nice little summary of all of the different tasks that I've asked my agent to do.
And it can be useful when helping other people build out similar workflows.
I do not recommend making manual changes to these memories, right?
Just let the agent store the memories however it wants to because the agent has a skill that keeps these updated and it will get better over time.
The agents.md file on the other hand is your manual memory.
This is meant for you to keep it updated.
You can tell the agent to update it or you can manually update it if you want.
The third huge capability that Codex has is plugins.
And so plugins are basically just Codex connections.
They are installable, reusable bundles that connect the OpenAI Codex AI agent to external tools, apps, and workflows.
And this allows them to have other functionality and they can connect to all of the tools that you use.
And so if you go to the plugins tab, with two clicks I can add Slack and I can add Gmail.
Once you've added Gmail, and you can see here this plugin actually comes with some skills.
So there's a Gmail skill and an inbox triage skill that is part of this plugin.
Right now, I think there's over 100 different plugins that allow you to connect Codex to the tools that you already use.
So, let's go over an example.
So, I created a new project, which is a folder on my computer called brand deals, and I'm going to say, "I want you to look at my email from the past 2 weeks.
Find all the brands offering to do paid promotion, do research on all of them, and put them in a table with relevant notes." And then I can type the at sign, and here I can type in Gmail.
All of the plugins show up in this menu, so we can very easily at mention the plugins.
And this is a quick way to see which plugins that you have access to.
And I really like using this menu, and then we can go ahead and click send.
And now Codex will search through all of my recent emails because I've given it access via the plugin, and it will get back to us.
This is taking about like 5 minutes so far, and I have another idea for a task I want an AI agent to do.
So, I can very easily just hit this new message button, and I can type out exactly what I want.
So, I'm making my next video on CLIs and MCPs in Codex.
Please look at my Notion, and so I at mention Notion, which is a plugin, and find my high-quality long-form scripts.
Then I want you to please write a script on this topic.
Use Excalidraw diagrams.
This is a skill, which I'll get to next.
We can run this prompt.
And so now we have two AI agents working, both working with plugins.
So, the brand deals task is done.
If we go ahead and open up this document right here, we can see that it created file, and it went through all of my emails and wrote down the companies offering to sponsor my videos all the way down.
It literally went through all of them and took detailed notes and said, "What did they ask for?" And also they added some research notes.
And then I could ask it to respond to the emails because of the same way that it can read my emails, it can also send emails from my account.
And if we go to the Notion plugin that we tested, we can see here that it created this document, and it went through my Notion, found all the scripts that I made previously, and it said, "Do you ever feel like AI coding tools are getting more powerful every
week, but also somehow more confusing?"
And this is literally in my voice because it went through all of my scripts, and it created this document using the Notion plugin.
And that brings us to the fourth capability that Codex has, which are skills.
And you can think of skills on Codex as reusable workflow recipes or SOPs that your agent can use many times over and over again.
In Codex, if you go up to the plugins tab, you see that there's actually a subtab within plugins, which are skills.
And in the skills tab, these are similar to plugins.
They're more like instruction files that your agent will follow every time it wants to use one of these skills.
So, all your skills can be found in the plugins folder in the skills subfolder.
But, how do you actually create a skill?
In order to create a skill, you can type a prompt just like this.
I want you to create a skill called brand deal analysis that does blank.
And I could enter this prompt in and create this skill, and then once that's done, in order to use any skill, you just press this slash.
And so, I can use my Excalidraw diagram skill just by use my Excalidraw diagram skill just by typing slash.
I can also use my YouTube researcher skill or my YouTube thumbnail skill, which we'll get to in just a skill, which we'll get to in just a second.
These are all skills that I can quickly access by typing slash.
These are different than plugins, which I can access by typing at.
So, there's two main ways to create a skill in Codex.
One is the prompt to skill, which is basically just I want you to create a skill called blank that does blank.
This will give you a reusable instructions file, which is a skill, but I don't think it will be as good as option one.
The best way to create a skill, in my opinion, is to use the manual workflow method, where you basically ask your agent to do something, it'll do it, and then you basically go back and forth, right?
Tell it to create something, and say, "Hey, I want you to change it." And you say, "Get the output right." If you're not happy, you iterate more, and you kind of go back through this process, and you loop around until you're happy with the output.
And then you reverse engineer the skill, and you will basically end up saying something like, "I'm happy with this output.
Turn it into a skill." And I'll actually show you.
If we go back to Codex, and you can see here, I've asked the agent to put all those paid promos into a spreadsheet right here.
And I really like this spreadsheet.
I think it's high quality.
I think it's color-coded, right, by the different priority.
This is looking really good.
And remember, you can always full screen it to get a better look at whatever document it is that you're creating.
And so, well, all we need to do is we need to just type this in.
"I'm happy with this output.
Turn this into a skill that I can use." And the agent is smart enough to be like, "Okay, this is he wants to always be able to analyze his Gmail for brand sponsorships.
Alternate into a skill." And so, now it's done.
It says, "I've created and validated the reusable skill called brand deal researcher skill.md."
And we can open up this skill by uh clicking on the skill.md file, and we can actually read this over.
And it turned our entire workflow into these instructions.
And the AI is very smart enough at taking anything that you've done with your agent, and turning it into a skill.
So, now if we hit command N to create a new chat, "I want you uh to please do {slash} It's called uh to please do {slash} It's called brand deal researcher.
Right?
So, now we have this reusable skill that we can use at any point.
And you can see here it's using all of the same plugins.
It's using the Gmail plugin.
So, a skill can contain instructions to use a specific plugin because at the end of the day it's just instructions to do a simple task.
While our research brand deal task is loading, I actually went ahead and did a create Excalidraw graphics for the seven capabilities.
The video that we're doing now, remember this document that we created earlier which is this landing page that describes all of the different things that we're talking about in this video.
I went ahead and I just said, "Please look at that document and make a ton of graphics on a single Excalidraw doc.
Use the Excalidraw diagram skill and make the chart super useful." So, then the AI thought and worked for 4 minutes and 23 seconds and then it created this.
And we can open this up and we can just And we can open this up and we can just automatically basically import this project and boom, look at this.
It created this entire outline and I love this formatting.
This formatting that it did with the background here, I actually really like that.
And so, it created all of the diagrams.
Wow, this is actually fantastic.
It has the intro, all of the summaries, memory, plugins versus skills, skill creation.
Very cool.
So, now I can go back to Codex and I can say, "Wow, I really like the formatting that you did here.
Please update the skills so that you always put them in the containers that you did and I want you to always give me a link with all of the diagrams that you create when I ask you to create multiple diagrams just like this.
I love it."
And so, even once you've used a skill, maybe you've used it for a while and you like it and then you get surprised by something valuable like this diagram right here or these this formatting, you can always just say, "Hey, I want you to
always format it like this." And so, every time you use the skill is an opportunity to make the skill even better.
Just tell your agent to update the skill to always do something.
The next capability of Codex is built-in GPT image to access.
So, you can generate images with the best image model in the world directly inside Codex.
So, I created a new project called content, and here I'm saying, "Generate a product photo for my uh sweater company.
This is our first sweater.
I can paste an image of a sweater, and you can see here, this is the first sweater, knitted sweater that I'm making.
Please create five images with GPT/image/2 of different models of different of different models of different nationalities wearing this sweater.
Three of the images should have one person, one image should have three people, the last image should have five people."
And so, we can run this, and within the content folder, it's going to create images.
And those images will live inside the content folder once it's done.
And look at this.
You can see here that it's using the skill.md for image gen.
And so, this image gen skill is built into Codex.
And so, you can go to Codex, click on skills, and you can see which skills are built into the system.
And you don't have to enable them, but these are recommended skills by Codex, and one of them is image gen.
And by default, this uses the latest and best image model.
And you can see here that it generated one of the images.
We still need to wait for the other four.
Okay, here, it just generated the second image.
It just generated the third image.
The next image should have three people in one image.
And here we go.
Here is number four, and here is number five.
And now that you know that you can generate images, that brings me to the next capability of Codex, which is browser use and computer use.
And so in order to use computer use, it is a plugin.
So remember, you just type in at computer use.
Now I can say, "Please open the Canva app on my computer, make a new presentation, and put each one of these images on each of the slides." So one image per slide, there should be five total slides.
We can simply tell Codex to control our computer and it will do it.
And as you can see here, we see list Mac apps and now it's looking at Canva.
You can see here, it's controlling my computer.
I'm not even touching it.
It just created this presentation.
It's clicking in Canva.
Okay.
Oh, it just ran a script, I believe.
And so it just ran a script.
It's looking at Canva and you can see here it created this presentation and it has all five images on the computer.
And you can see here, this is the mouse.
You can see the mouse outline of the computer use.
It will move around from time to time.
It created this presentation for me.
So this right here is a browser and you can see that if we literally exit out of this and hit plus sign, we can see browser.
And when we open this index.html file, it's literally opening this in a browser.
So now I'm going to say, "Turn this into an app, then test the interface and make sure the buttons and navigation work using at browser and navigation work using at browser use."
And I can run this prompt here.
And now it's going to take this, turn it into an app, and then it's just going to test it by controlling the browser inside Codex.
Okay, you can see here, it's kind of like a different color.
Look, this is my mouse right here.
This is the browser use.
So the AI is actually controlling the browser and it's going to start testing them.
So, we see that it's test start button, and the AI is now taking control of the screen.
The first visible check showed the app controls, but the start button smooth scroll is too subtle for live verification.
You can see here it's moving around.
It's scrolling through the app.
It's clicking on the different quizzes that it's created.
Now, it's testing quiz completion, and now it just marked this a correct answer.
So, it's making sure that all parts of the application work.
Oh, now it tested the side panel, and now we're at the bottom.
This is very cool.
Anything you can open up in the browser, you can ask browser use.
You can use the @browseruse plugin to test it directly inside Codex.
And finally, capability number seven inside Codex is automation.
You can create automation simply by talking to Codex.
So, remember this Excel sheet that we created in this chat here in the brand deals project called research paid promo brand deals.
We created this spreadsheet after it analyzed all of my incoming emails with the email plugin.
Then, we took this output, and we turned it into a skill.
I'm happy with this output.
Please turn it into a skill.
Now, what I can do is I can repeatedly do this on a weekly or daily or hourly basis.
I can simply say, "Please do this every single Friday at 9:00 a.m., please, and update this sheet."
What this is going to do is it's going to create an automation.
Notice here we only see two automations.
Here it says, "I'll set this up as a recurring automation for Fridays at 9:00 recurring automation for Fridays at 9:00 a.m."
And we can see it directly in the UI.
Created weekly brand deal sheet updated and if we click on this, we can see the automation details.
We can see that at Friday at 9:00 a.m.
We can see that it's active.
We can see that the last run is empty because it's never run before and we can see that we are using the brand deal researcher skill to scan Gmail for paid promotion sponsorship brand deal influencer campaigns, etc.
So created this and now if we go to the automations tab, we now see it up here.
We can see Friday at 9:00 a.m.
We have another one at Fridays at 9:00 a.m.
We can click on this and you can edit it directly.
And finally, we have a bonus capability inside Codex which is something called Chronicle.
So this is a brand new feature within Codex.
If we go to settings and we click settings and we go to personalization and we scroll down, you will see an option to to use Chronicle research preview.
When you turn this on, what this will do is it'll say that it is running and this one may seem a little bit invasive.
What this one does is it basically is always recording your screen so it has context with what you're working on.
And so if I go to my presentation here and I just kind of scroll around in my browser, use kind of scroll around in my browser, use Chronicle.
Uh tell me what I should add to my Codex visual presentation.
And so it shows you here all use Chronicle skill for live screen work content then cross everything over.
Chronicle's available so I'm checking the fresh screen state first and look at this, you can see that it's pulling up recent screenshots of my screen here.
And so it is pulling old images and it's getting context from what I've had open on my screen recently.
And so it's able to keep track of all the things that I'm working on and the way that you invoke it is just say use Chronicle.
And so it says based on Chronicle, I'd add these to your visual presentation.
A Codex super app map, a full control loop diagram for prompt files, a Chronicle demo slide, reusable skills slide, and why Codex over other tools slide.
And so, it automatically, and I didn't give it any [clears throat] information, I didn't upload an image or give it an Excalidraw document, it just knew what was in this slide deck here.
And it's able to understand it because this Chronicle feature watches my screen and takes screenshots constantly.
And so, those are the main features and abilities within Codex.
And in summary, we covered that Codex has full file control, meaning you can create, edit, and delete any file.
It has persistent memory, and whether that is automatic memory or manual memory in your agents.md file.
We also talked about plugins, which are built into the platform.
It connects your agent to all of the tools that exist.
You can create skills by taking any task that you create with the agent and then converting that into a skill.
Your Codex agent has access to image generation, and it is the best image generation model in the world, which is generation model in the world, which is GPT-image-2.
We also have browser control and computer use, and this allows your agent to control your computer like a human.
The human interface, which is a keyboard, a mouse, and a screen, can now be used by agents, and it's getting better exponentially.
And we also have automations.
You can take any useful task that your agent does and say, "Hey, do this every day.
Hey, do this every hour.
Hey, do this every week.
Do this every month." That type of thing.
And then as a bonus, we talked about Chronicle.
Anyway, thank you guys so much for watching.
I hope you got a lot out of this video.
I make videos like this every single week, so make sure to subscribe and hit that like button.
It helps me out a lot.
I'll see you here for the next video.