全网最全，50分钟全面掌握新版Codex～【附完整文档】

Codex has updated a lot of things, so Codex's new version of the full strategy is here. It can build more sites and even develop its own plugins. Pictures can

is here. It can build more sites and even develop its own plugins. Pictures can

be directly copied and modified, and a command can create multiple independent conversations to carry out different tasks at the same time. How do you use so many functions? In

this video, I will take you through a few real cases to learn about the basic use of Codex and some details of the update. Let's see how it works in office and deep learning Then enter the long-term memory, skill, MCP and automation These advanced play methods Finally, talk about the hook and hook in the AI programming system

Distribution, Worktree and Subagent This video will be a little long I have compiled the relevant notes and keywords into a document Everyone likes and follows, let's start directly OK, let's talk about CodeX. OpenAI has integrated CodeX and ChatGPT into the same application. The update is very fast and multi-intelligence is more stable. You

same application. The update is very fast and multi-intelligence is more stable. You

can directly control the voice, create, view and adjust CodeX tasks. There are

many plugins, and you can even create your own plugins. Now the picture can be added to the view to be modified. And DeepSick has also released a method to integrate Codex. I wrote this in the article. After talking about so much, how to

integrate Codex. I wrote this in the article. After talking about so much, how to use it specifically? Let's go down. Open TrackGPT, the upper left corner is the core function area, the lower left corner is the project area. The top can be switched between TrackGPT and CodeX. TrackGPT includes chat and office functions, while CodeX is

more suitable for code development and project execution. The work in TrackGPT is basically the same as CodeX, but CodeX has more functions, so I will use CodeX to demonstrate. Let's go to the AI dialogue box in the middle. Let's solve the

to demonstrate. Let's go to the AI dialogue box in the middle. Let's solve the first question first. There are so many models here, when should we use what model?

Generally, for development or more difficult tasks, we can use GPT-5.6 SOD. For smaller tasks, we can use Tera. Actually, Luna is used very little. Click the plus sign next to the project on the left to create a project. This is also the core of the desktop AI Agent folder. How to use this folder? You can

put some documents, tables, pictures, codes and task rules in the folder. Codex can

not only read and analyze the content, but also create, modify and organize documents directly. When creating a project, we can add multiple original folders. The main folder

directly. When creating a project, we can add multiple original folders. The main folder here is the default workspace. This folder usually contains some task rules and the final result. Other folders are for adding information, such as brand regulations, historical plans and reference data, or your own local knowledge base. Codex can

search and call data from this folder. You can understand it as an additional database. When needed, Codex can also directly operate the folder inside. For example, I

database. When needed, Codex can also directly operate the folder inside. For example, I usually organize the paper, technical report, and summary notes of the model. This is

used as my supplement folder. After DeepSig Harness came out recently, I need to study how to use it. I will build a DeepSig Harness study folder and set it as the main folder. OK, I will send the first prompt to Codex Help me deep dive into DeepSeq Harness Collect official articles, project

warehouse and all release instructions Search for representative articles, video and test content on the platform Then analyze the operation methods of different authors and make horizontal comparisons Finally, send me all the links and labels When the first time Codex is executed, I usually set the model to high, hoping that it will do better the first

time. The hint will also be written more clearly. Click the plus sign in the

time. The hint will also be written more clearly. Click the plus sign in the left corner and select the planning mode. In this mode, it will generate a planning plan first, and then execute it after you confirm it. There is also a target mode here, which we will talk about later. Send a prompt, AI will load the tool file and search the web page to confirm the GitHub warehouse and ask a

few questions. For example, he asked me how much the video needs to be selected.

few questions. For example, he asked me how much the video needs to be selected.

I choose this one. The link and the decoder. Ask me what size of selection I need in my article. Then I choose the one he recommended. I also choose the one he recommended for this question. After confirming, it will generate a plan. We

can see if it meets our requirements. If it does not meet, we can modify it. Here I will let it execute directly. This execution time may be relatively long.

it. Here I will let it execute directly. This execution time may be relatively long.

Let's look at its results directly. It finally output a lot of content, including in-depth research reports, Redmi files at the entry of the party, official data source, selected article source, and 13 analysis cards for different authors. The small card in the upper right shows the planned document, the output result, and the source link, such as the GitHub

address of this DeepSeq Harness, which allows us to quickly understand the code. Click on

this in-depth survey report, you can see the specific content in the Codex display area on the right. The summary is very good, there are core frames and the agent execution track are very good. There is a link to the selected article below, press the command key and then click it, you can see the page

on the right. So we can learn directly in the internal browser of Codex, it is very convenient. If you think the report is not very clear in some places, you can directly select the pieces that need to be edited and add them to the AI dialogue box in the middle. For example, the number of inputs here

can be asked directly to the AI. There is also a voice input function on the right side. Click "Settings" "Voice" Here, you can set up your long press button to press the "speak directly" button and then convert it into a hint. What does

this mean? You explain it again.

Another point, for example, here to modify the full line and the sand box, we can directly choose it, click Edit, Send Requirements, AI will help you to write the content, and then you can choose to refuse or accept, here choose to accept. Click

on the file icon in the upper right corner, you can see the content and the structure of the file folder, this sidebar is really convenient. Let's go back to the AI dialogue box in the middle. You can see that there are three options for all lines: "Ask for approval", "Help me approve" and "Complete the full line visit".

In general, we choose the second one, "Help me approve". "Ask for approval" means that every time you face a risky operation, you will ask if you can do it.

"Help me approve" means that AI will help you judge in advance. The low-risk operation is handled by AI itself. High risk to let you confirm complete full-time visit is full-time AI in general is not recommended Let's click the little icon on the side again to open the side chat, which is a temporary chat. In this chat box,

it will not affect or execute the current operation. Of course, if you have some ideas for天马行空, but just thinking, you can open this sidebar, here to plan a plan, how to write a hint. For example, I would like to ask what is the difference between Deepseek Harness and PIE, because I think they are both very simple agents.

AI gave the result and loaded the official instructions of Pi and the official warehouse of DeepSick Harness. He explained that Pi is like a light Swiss military knife, and DeepSick Harness is a workbench that can constantly install and set up.

DeepSick Harness's all-in-one key is really strong. Next, I want to study the two models, DeepSync V4 Pro and Flash. Is there a big improvement in using DeepSync Harness?

I have simply tested these two models before, and the performance of the codecs was included. There are relevant evaluation materials in my supplement folder, so I input

was included. There are relevant evaluation materials in my supplement folder, so I input the prompt. For reference, I compared the test screen of the DeepSync V4

the prompt. For reference, I compared the test screen of the DeepSync V4 Pro and Flash to the test screen of the DeepSync Harness. The sender

indeed found the complete test material of the DeepSync V4 Pro and Flash.

Then the library and the input of the accessories have also been generated.

There are six sets of old and six sets of new uses. OK,

it's ready to be generated. Give me a Markdown file with some keywords and rating standards. For me, I often use this supplement folder as my supplement knowledge library. When studying DeepSeq Harness, it added a track function. If I want to read the content on the page, I can use the quick photo function of

Codex. At the same time, press two command keys to send the page quick photo

Codex. At the same time, press two command keys to send the page quick photo directly to AI. Let it explain the purpose and specific view of the page. Click

on this page. Click on this to view the text It will bring the URL link in the quick photo This allows AI to see the content on the desktop very clearly This is much better than ordinary screenshots AI will soon see the content of the page And list the content representative meaning Time line, button function and bottom

performance indicators On some pages that are not easy to understand It is recommended to use this quick photo function After the survey is completed, we can add this content to the Markdown article in the depth survey general report. We send a prompt, it will list the files to read the article of the depth survey general report. At

this time, for example, we just wrote the prompt, and we want to say how to see this page to be written more clearly. Click "Send" There will be a adjustment direction here, which is to submit the "Deposit Interrupted" model operation directly to it.

Add it to the current operation of the top and bottom text If you don't click it, wait until this round of dialogue is finished and then send it to AI. Then we click this adjustment direction. OK, now it has added this

AI. Then we click this adjustment direction. OK, now it has added this thing Added the location of the six areas of the page The specific usage of each button and the contrast content are very good We enter a diagonal bar again Click state You can see how much the top and bottom lines are used Your

drawing ID and your amount Enter a diagonal bar and a compression You can compress the top and bottom lines When it accounts for about 60% I suggest you to compress it. After these articles are reviewed and generated, I need to copy them

compress it. After these articles are reviewed and generated, I need to copy them to the supplement folder, which is in my knowledge library. In the default main folder, I will enter the prompt. I have written two articles, namely the depth review report and the evaluation article. Please copy these two articles. and collect the

core information to the AI model's supplementary folder. Codex will locate the two articles and the supplementary folder, sort and record, and then file it. In

the end, it will generate a file file. Because my supplementary folder is managed by Obsidian, it will also add some corresponding notebook properties. This is

very useful for my knowledge library summary. I don't know what you think about this.

The reason I'm going to do this is because when I use the AI file folder for a long time, the file folder will become very messy. Maybe I ask a question, ask a question, make a conversation here, and then open another one over there. and then we have to render it again. So this file folder is easy

there. and then we have to render it again. So this file folder is easy to mess up. I don't know if you have the same problem as me. If

you have, please press 1. Click on the icon in the upper right corner to open the middle of the bottom. At this time, we can enter Cloud. Two AI

can be used together. You know what I mean. Back off. I'm going to start putting in B. By the way, you can set up a pet in the settings yourself. When communicating with AI, it will interact with you and lead a companion. I

yourself. When communicating with AI, it will interact with you and lead a companion. I

saw someone on the Internet who made it into their own cat and dog. It

looks particularly fun. You can go to the dog. OK, we have learned the basic interface and usage of Codex through the case. But in the real office scene, we are more faced with local office files such as Excel, PPT, Word. So let's take a look at the codex version. local file processing capabilities. Now I have a mobile

phone in this folder. The sales data for the past few months include the order date, sales channel, product style, sales volume, sales volume and refund situation. Now the boss asks me to do a business report based on this data. In this process, we need to do data analysis, draw a chart, make a presentation, and maybe prepare a

speech. Let's send the first prompt to Codex. Currently, there is a mobile phone store

speech. Let's send the first prompt to Codex. Currently, there is a mobile phone store in the folder with sales data for nearly four months. Please make a business report based on this data and finally produce an Excel table. What is the overall framework?

Which one is better? Which one is worse? How to analyze the style? Analyze each

channel and style separately. After it's generated, we can add a few more little discoveries. We can add a spreadsheet. Click it, and it will provide us with some templates. Let's choose the second one. Then we use the data analytics plugin.

one. Then we use the data analytics plugin.

It will follow the model I have designated to retain the visual system. First, it

will load the data analytics plugin function, and ask it to execute it according to the specific analysis process. First, the data quality test, the report construction, the chart production, and finally the result verification, and emphasize the avoidance of directly using empty Excel tables.

After he confirmed it, he started to look at the table After the data quality test is passed Then he will operate according to the skills of Spirit Sheet Finally, he generated a table The effect is very good The style is also good The following contains four content analysis Of course, we can now modify

it directly in the table. Select a part, for example, I think it needs to be more detailed here. Select this section of content and send the demand to the AI dialogue box for modification. At the same time, you can also click the WPS icon in the upper right corner to open and edit through WPS. Codex is very

good at interface switching. Next, let's take a look at Word. We paste the prompt.

This time we use the "Documents" plugin. It also has some text templates. What if

you want to use your own template? Right-click on "Create Template" and it will give you the corresponding prompt. Let's create a new dialogue to create a template.

This is a template I prepared before for e-commerce management analysis report. We let it create based on this template. You can directly click on the send prompt. It will

use Documents to read and understand reference documents in order. Then use Template Creator to generate a template. The template will be generated soon. Back to the conversation just now, the insertion of @Documents again. The first option is the template we just created. Click

on it, use it, and then send a prompt. Based on the Excel form just created, help me create a Word version of the business report. Wait a minute The generated file and the just made model are basically the same Fully meet our requirements In this Word file You can also select a certain section to modify And send

it to the AI dialogue box Very convenient Next is ppt, also send a prompt, then use the presentation plugin. Here I don't choose the model, let it directly based on word to train e-commerce management report, generate ppt. It

will also load the presentation plugin, although the final result is not particularly good. But if you add some skill skills or molding, the effect will be better.

good. But if you add some skill skills or molding, the effect will be better.

Codex is more advanced in the modification of ppt. You can directly add a part of the view by commenting. This is really good. If you want to add some more pictures on ppt, you can also directly use the deep picture function of codex to add. Very convenient. The last one is HTML. We send a prompt,

to add. Very convenient. The last one is HTML. We send a prompt, based on the business report, generate an HTML web page, write the requirements, The result is not bad, and it also gives some summary and suggestions. For the

Html edit, you can also click on the Pinch button in the upper right corner, select the part that needs to be pinched, and add comments to edit directly. In

addition, click on the small icon on the left, and you can also directly edit the color, font and size. This is very convenient and convenient for developers. OK, let's

talk about something else. I woke up in the morning and suddenly found a lot of things in my hand. I want to look at e-commerce data, analyze the recent financial situation, and study the recent Deep Seek Harness. At this time, we can write down the needs, let Codex automatically help us build missions, and then work separately. We

can write such a prompt. The point is to clarify two missions.

帮我安排两个任务。 第一个任务,找到DeepSickHarness研究文件夹,

在这个文件夹里新建一个任务 研究DeepSickHarness是什么,怎么用 有哪些优点,缺点 最后整理成一个研究笔记。

第二个任务,找到电商数据的这个文件夹, 在这个文件夹里新建一个任务, Analyze the e-commerce data inside Find the latest sales situation and existing problems Finally, generate an analysis report And give a few practical suggestions Both tasks are stored in their own folders Come back and tell me after completion What conclusion did each task come up with?

Where is the file stored? After sending, Codex will find the corresponding folder, and build two new task conversations. One to study DeepSeq Harness, and the other to analyze e-commerce data. Here it automatically writes the prompt, and shows that another task is sent by

data. Here it automatically writes the prompt, and shows that another task is sent by TrackGPT. These two tasks can be executed at the same time. After they are done,

TrackGPT. These two tasks can be executed at the same time. After they are done, Codex will come back to report in unison what each task has done, what conclusions it has made, and where the generated files are placed. OK, let's see the results of these two tasks. You can open the corresponding chat below to see the corresponding

folder. It's wonderful. Oh my god. Here I also talk about the fact that sometimes

folder. It's wonderful. Oh my god. Here I also talk about the fact that sometimes it takes some time for AI to complete a task. We don't have to wait all the time. We can create a new conversation directly and continue to arrange other tasks. Let multiple conversations be carried out at the same time and promoted simultaneously. This

tasks. Let multiple conversations be carried out at the same time and promoted simultaneously. This

will save a lot of time and efficiency. OK, we have learned about the local file operation capabilities of Codex and some of the skills used in office files. We

also used some plugins. Next, let's take a look at another big function of Codex, "Drawing". Now there is a product picture of an outdoor electronic watch and a package

"Drawing". Now there is a product picture of an outdoor electronic watch and a package material picture in this folder. We can directly let it draw a picture. I have

written the prompt clearly. I am a seller of Amazon's North American store, selling outdoor electronic watches. The folder contains these pictures. It is based on the picture specifications and

electronic watches. The folder contains these pictures. It is based on the picture specifications and machine vision recognition logic of Amazon North American store. Make 5 pictures of 2000x2000 pixels.

In addition, make 7 pictures of A+ and write the necessary time to search online.

We send a prompt. You can see AI first checked the six product images in the file, and tested the requirements of Amazon's current public image, and then produced it.

In the end, it can be seen that it indeed searched the Amazon image standard and product image prompt online, The picture is generated The effect is also quite good But sometimes the generated picture does not meet our requirements We need to modify the picture Codex right-hand side updated the function of picture editing You can modify the picture

directly on top Click here to add comments For example, add a "increase rainy day scene" Send this prompt again It can send the coordinates of the comment point to Codex for deep painting Of course, there is also the function of cleaning and adjusting the size But in fact, many detailed image modifications We can't describe the

prompt clearly At this time, we can use GitHub on a free plugin called Covert Here you can directly send this command to Codex for installation I have installed it here, so I won't show it We just need to drag the picture into the picture cloth and we can mark it. For example, this poster picture only

needs to be dragged into the picture cloth to mark the content to be modified.

Here I circle out the numbers on the surface of the watch, say it will be changed to 9.14, and then circle out the sunlight on the left, say remove the sun. Then, we will give this screenshot to Codex and then add the poster

the sun. Then, we will give this screenshot to Codex and then add the poster we just created to it. Let it be modified according to the screenshot. This way,

we can describe the image in the picture more accurately. OK, it's generated. It's actually

pretty good. Many times, we can use this method to accurately describe the prompt to adjust the AI deepening. Finally, I will tell you a little tip. When we make a picture, we can actually communicate more with AI. For example, this picture of a child's short sleeve, from a product perspective, its selling point is to absorb the dryness.

We may write directly after making the picture, but after communicating with AI, it will tell you that it is more suitable to write on the page of the photo to quickly dry after exercise. It's not easy to catch a cold. This will make your mother feel like a giant when she buys clothes. So when you work, you should really chat with AI more. You say yes. The previous content basically covers 70%

of the codex usage. Are you still there? If you think the information is a bit big and a bit tired, you can click on the double attention first, rest, and then come back to see it. Then we will move on to advanced play.

First, let's take a look at Codex's long-term memory. We know that Cloud Code has Cloud.md, and Agents.md corresponds to Codex. Its configuration is mainly divided into two

has Cloud.md, and Agents.md corresponds to Codex. Its configuration is mainly divided into two levels, and the priority classes are different. The first is the full configuration, the path is in the system layer, click on the setting, personalization, and here is Agents.md. Usually we will write personal number, code style, and common tools here. For

Agents.md. Usually we will write personal number, code style, and common tools here. For

example, I wrote here that China is the first priority and before performing destructive operations, it must first show the specific plan, mark which operations are irreversible, and then execute after clarifying. This is seen in the article. The second is the project-level configuration. The

after clarifying. This is seen in the article. The second is the project-level configuration. The

path is usually the root path in the current work area, mainly used to provide the top and bottom text of specific projects. such as the technical requirements of the current project, the structure instructions, etc. On the priority level, the project level must be greater than the overall level. That is to say, the closer to the current work

path, the higher the priority level, Codex will automatically merge these layers and follow the logic of the latter covering the former. Let me give you a specific example. Look

at the structure of this folder We put a general agent's dmd at the top It says the general rules and code style of the front and back end In the front end folder, a separate front-end exclusive agent's dmd is placed In the same way, a back-end file folder also has a back-end rule that stipulates the content of

the API design. In this way, when you work under the front end, it will first read the top-level general rule, then use the front-end exclusive rule to cover it, layer by layer, and the logic is clear and it will not be messed up at all. In the previous example, we have seen that Codex

at all. In the previous example, we have seen that Codex can directly operate the computer terminal. This ability is very useful because many software now provides its own CUI. For example, Dingding CUI, Qiwei CUI. After installation, we can let Codex call it in the terminal to further link the various functions

in Dingding. Similar tools and GitHub CUI can be used to operate warehouses, I

in Dingding. Similar tools and GitHub CUI can be used to operate warehouses, I use my most common facial CR to demonstrate. If you install it, you only need to send this question. I have a tutorial article about Git and GitHub here. I added some comments. We can send this

article to Codex Let him read the text first Then review it again Then modify our content according to the meaning of the comment For convenience I will let him mark the contents of the AI modified into red So you can see where the AI has modified Send a prompt Codex will call for the skill of modifying non-fiction

articles To read the comments inside OK, he finally read my comments. The first comment is to make it into a picture. The second comment is to modify a section of content. Let's take a look at this picture. It's actually pretty detailed. Very good.

of content. Let's take a look at this picture. It's actually pretty detailed. Very good.

The red font at the back is the AI modified content. If I think there is no problem, I will change this content to black. This is very useful for me. Just add comments. Then AI can help me edit the content. All my tutorials

me. Just add comments. Then AI can help me edit the content. All my tutorials are written like this. In the previous video, we demonstrated Codex editing articles through non-Library CLI. Next, let's take a more realistic example. Let's say I'm a computer installation studio.

CLI. Next, let's take a more realistic example. Let's say I'm a computer installation studio.

Customers may order from different platforms, such as a certain treasure, a certain east, a certain small shop. They may also find us through private channels or offline stores to configure computers. When the customer orders, in addition to the order number, price and contact

configure computers. When the customer orders, in addition to the order number, price and contact information, they will also provide some installation information, such as whether this computer is used for editing or playing games, and specific configuration requirements such as CPU, graphics card, memory, etc. These businesses are mainly involved in two departments: Sales Department is responsible for receiving

orders and confirming needs, and the Technology Installation Department is responsible for reviewing, configuration, installation and delivery. At this time, we can use the non-digital multi-spec table to communicate information.

and delivery. At this time, we can use the non-digital multi-spec table to communicate information.

I have already created a non-bookmarked form and an Excel form here. Record the data of the orders from each platform. We only need to send a prompt to help me fill in this Excel form into this non-bookmarked form and attach this non-bookmarked form link. Codex will read the text and order data in Excel, automatically skip the repetition,

link. Codex will read the text and order data in Excel, automatically skip the repetition, and then call the non-coding skill to execute the terminal command. OK, now the data is filled in. It tells us that 12 data have been successfully imported and 0 have been repeated. We click on the multi-tab. You can see that the data is

very clear. Information such as CPU, motherboard, and graphics card has been automatically filled in.

very clear. Information such as CPU, motherboard, and graphics card has been automatically filled in.

Finally, we can mark the current communication, installation, delivery, cancel, or complete in the order status column. This method is very suitable for inter-department information and business communication.

Of course, we can also let AI continue to analyze this table to calculate which CPU, graphics card, and memory models require the most. to provide reference for subsequent purchases and back-up. Similar methods can also be applied to e-commerce product surveys, customer follow-up form, sales, clues management, etc. In the scene, the data that was originally scattered

is actually flowing. OK, we used Codex to connect some operations executed by the non-finalist CUI. Next, let's take a look at the plugin. Simply put, the plugin is a

CUI. Next, let's take a look at the plugin. Simply put, the plugin is a collection of Skill, MCP, and CUI. The official package of the things needed to do a kind of work. We click on the plugin in the upper left corner. You

can see that there are a lot of plugins here. In the upper right corner, there is also the creative plugin. That is to say, we can develop one ourselves.

So many plugins, I don't want to talk about them. We focus on choosing some representative ones. More introduction in the article. First, video type.

representative ones. More introduction in the article. First, video type.

Let's talk about Remotion and HyperFrames. Both use code to make videos. Remotion uses React, HyperFrames uses HTML and CSS.

make videos. Remotion uses React, HyperFrames uses HTML and CSS.

For example, I created these images with these two X keys. How to make a good animation in the following skill is also thought together. And then there is this chat card X key. It is an AI editing X key for dialogue. You can

directly let it cut out the sound, delete the sound, press pause, add subtitles, and even add some key animation. This plugin may not be available in the plugin market.

You need to run this prompt to complete the installation. I use it myself. TraderCard's

cut mouthpiece is really smooth. I still recommend everyone to use it. But this issue I will not specifically expand. Because the whole process allows AI to cut a mouthpiece.

The first part will be combined with skills, scripts and analysis. Finally, I will give the editing to Chatka. There are many contents in the whole process. If you are interested, I can do a separate episode of AI digital voiceover and AI editing. I

don't know if you want to watch it. The second is the half-work type of plug-in, like doing PPT, doing text files, which we have used earlier. I won't repeat it here. If you need it, just go to the plug-in market and install it.

it here. If you need it, just go to the plug-in market and install it.

The third one is about this. We know that Cloud has its own Cloud Design.

There is actually an Open Design. OpenAI is the product design. Let's take a look at the demonstration. Let's let AI do a personal main business. The style

is as simple as OpenAI official website. Then we add the product design. It

can create your own model, but I have not created it. You can create your own model and do the design. OK, send a prompt. It will call this product design plugin to determine the visual direction and page structure. Will give me three visible directions. It even called my computer here to take a look at the OpenAI

visible directions. It even called my computer here to take a look at the OpenAI official website. Visually, I borrowed the black and white and large font version on the

official website. Visually, I borrowed the black and white and large font version on the OpenAI official website. This is very good. OK, Codex asks us to fill in some information. I'll send it to him after I fill it in. He will use the

information. I'll send it to him after I fill it in. He will use the deep drawing function to generate the front page first. Here he gave me three solutions.

I think the first one is probably the best looking. But I don't need this head. I can tell him this is good, but don't want this head.

head. I can tell him this is good, but don't want this head.

Here he is when developing the load multiple agent, this is the main point to talk about the self-intellect, generally for some more complex tasks, it will automatically load multiple intelligence to do work, that sometimes you want it to load multiple intelligence to work, you can write in the prompt, arrange what intelligence to do, arrange another intelligence to

do what, so that AI will arrange the self-intellect to work for you, Now it says it needs my authorization to preview locally. Then we send permission to preview locally.

AI checks if the interaction between the desktop and the mobile end is consistent. Then

it modified it. In the end, it gave us a very good effect. I think

this design is really very similar to the OpenAI official website. Very consistent with the requirements I wrote in the prompt. Nice. In the development process, I also recommend using the SuperPower plugin to develop. It will first clarify your needs, and I will write a test plan and then manually do it. After doing it, it will automatically review

it. Finally, let's talk about the browser plugin, computer use and recording skills plugin. First,

it. Finally, let's talk about the browser plugin, computer use and recording skills plugin. First,

let's talk about Chrome automation in the browser interface. The official showed some wonderful cases. Codex can automatically browse the OpenAI developer community forum, and grab the related

cases. Codex can automatically browse the OpenAI developer community forum, and grab the related posts of Codex in the last week. Summarize the theme, key questions, user emotions, and generate a structured Excel table, and verify the data. The second case is to find out the relevant email and local PDF data from the mailbox, automatically match

the date, amount, and business information, fill in the reimbursement system, upload the copy, select the category, and submit the form. The third example is to let four agents play the same online reading game in four independent Chrome labels. Codex will create a game space first, and then all the agents will join at the same time. According to

the same prompt, the final creation of four different lighthouse paintings will show the ability of multi-agent parallel and multi-agent coordination. In addition to this Chrome automatic insert, there is also a chat GPT built-in browser. We add the insert of this built-in browser. Let

it fill in the MBTI test Codex is automatically opened on the right side of the browser to fill in The final result is INTJ What is the difference between Chrome Autonomous and Codex built-in browser?

In a word, Chrome Autonomous controls the Chrome you are using The account has been logged in The web page has been opened It can be operated directly, such as sorting data from the background, filling out the corporate reimbursement system, handling the background of e-commerce merchants, or opening multiple labels at the same

time, allowing multiple agents to work together. Codex The built-in browser is directly on the right side to open the web page, suitable for temporary information checking, filling out questions, testing the web page, and running the registration process. Open it and you can use it. The operation process is also clearer. Then the computer use plugin also looks at

it. The operation process is also clearer. Then the computer use plugin also looks at the official case. In the first demonstration, Codex automatically controls the environment of the plug-in on Mac. It not only self-edits and runs the app, and run the game like

on Mac. It not only self-edits and runs the app, and run the game like a real person. After discovering the AI opponent's bug of breaking the rule by two steps, he directly modified the original code in the background, fixed the problem, and completed the re-test. Then in the second demonstration, Codex automatically opened the system-based drawing software after

the re-test. Then in the second demonstration, Codex automatically opened the system-based drawing software after receiving instructions on the Windows system and directly controlled the mouse and the light to draw a goblin on the screen. These two cases show the system-level GUI interface and automated wall-mounting capabilities of the Codex cross-platform system. So, computer use is mainly used to

operate desktop tasks that can only be done by mouse and keyboard. However, it consumes a lot of money, takes a long time to execute, and will occupy your computer.

I will use less. Finally, the recording skill. In this demonstration, the user first let Codex open the recording learning mode, and then he manually demonstrated the complete workflow of uploading videos to the backstage of the museum, including the corresponding title and introduction from the table, uploading the cover, adding subtitle files, and setting up the thought. After Kodak's

full-time observation, not only did they understand the operation, but also automatically summed up this process into a dedicated skill pack that can be called at any time. When the

next video is posted, the user only needs to throw a file pack and give a command. Kodak can automatically run all the upload and fill actions in the background

a command. Kodak can automatically run all the upload and fill actions in the background according to the number of PN shown by the user before. We can actually post videos or fill in some forms, the recording performance is very good. That's all about the plugin. Next, let's take a look at SKILL. Skill is a skill for

the plugin. Next, let's take a look at SKILL. Skill is a skill for AI, such as AI daily news skill, skill for front-end UI design, skill for beautifying ppt. I also recommend a few skills here. The first one is the

beautifying ppt. I also recommend a few skills here. The first one is the find skill. You can find the skill through natural language. This gives you a

find skill. You can find the skill through natural language. This gives you a human-oriented skill, and this deep research skill. Let's make an AI animation skill that I use more. At the beginning, I will put some of the pictures or animation screenshots I have made in the folder as a reference for AI. I sent such

a prompt. This folder is my animation picture reference. Please help me create a skill.

a prompt. This folder is my animation picture reference. Please help me create a skill.

As long as I give a script, I can generate the corresponding picture. After I

confirm the picture is 55, I will call the hyperframe plugin to generate an animation video. Enter $ and you can choose the skill. You can see the system-based skill

video. Enter $ and you can choose the skill. You can see the system-based skill "Skill Creator" is installed without any manual installation. At the beginning, it won't create this skill directly, but ask me a few questions first. For example, it asks me what kind of production method I want to use, because choosing A may consume more money.

and there is not much room for later modifications, so I chose B, which is the form of AI-generated images. At the same time, I also proposed that I need to generate two versions for reference in the first stage. OK, he raised some questions again, and we can answer them in practice. This process is to continuously ask questions, confirm details, and wait for the final plan to be set, and then he will

execute the operation according to the confirmed plan. In the end, this skill was initially created. After the initial completion, let's test it. I sent him a script prepared and

created. After the initial completion, let's test it. I sent him a script prepared and gave him a instruction. This is my script. Please download the skill just created to help me generate the picture. Very fast, it does generate two pictures, but I found that the effect of these two pictures is not ideal. The picture is too messy,

and there are too many text on it. So I replied to him that these two pictures do not meet my requirements. First, I want the font to be black with black. Second, the whole page needs to be more concise. Remove the text at

with black. Second, the whole page needs to be more concise. Remove the text at the top and bottom. It was generated again, but this time I felt too monotonous.

Tell it that it's a bit too simple now I still want to use the form of a card frame This will make the visual effect better After adjustment This time the picture is very in line with my requirements After confirming the error I directly sent it the last instruction Please use the hyperframe to insert the video card

to appear in the order of the script Finally, it used the HyperFrame plugin and rendered the picture into a video Although the rendering process is long, the final video effect is still very good You can take a look at this greatly saved me the animation time After we achieve the effect we really need, we can modify the

skill we just created. Send a prompt, "Fulfill skill according to the content I just created." AI will automatically call the skill create function to modify the skill content. OK,

created." AI will automatically call the skill create function to modify the skill content. OK,

now it has updated and improved my skill. For example, it has updated the "Must use Deyee Black" font. This skill is based on the changes we made in the previous text, and added the problems we encountered when using it. So, every time we use it, we can write our experiences or the pitfalls we encountered into this skill,

so that it can fit your needs more and more. What we just did is to organize a complex process and our own experience into a skill that can be used repeatedly. For example, teachers can do a backup skill, e-commerce can do a product

used repeatedly. For example, teachers can do a backup skill, e-commerce can do a product update skill, self-media can do a script and division skill, office scenarios can also do weekly reports, meeting notes, or data analysis skill. Although there are many online skills that can be downloaded directly, the real skills are often made according to their own workflow.

MCP is no longer used, so I'll just briefly talk about it. Click the insert button, then click the small icon in the upper right corner to see the MCP page. Click it, then click Add Server, and you can configure your own MCP here.

page. Click it, then click Add Server, and you can configure your own MCP here.

Of course, we can also let AI help us connect to MCP. For example, I want to connect to a notebook LLM MCP, I can send this prompt, and AI will automatically connect it for me. OK, we have finished the skill, MCP, CUI, and plug-in part. Next, let's talk about automation. Why do we need to talk about automation

plug-in part. Next, let's talk about automation. Why do we need to talk about automation at the end? Because automation will integrate all the content we talked about before, including skill skills, X keys, MCP, etc. The concept of automation is actually very simple. It

is to set the work that was originally needed to be sent manually to be automatically executed by the fixed time. For example, we click on the "Package" in the upper left corner to arrange. Here will provide some recommended regular tasks such as daily reports, weekly review, and follow-up monitoring, etc. We can also click on the create button

directly The system will help us write the prompt We just need to tell it Which content needs to be uploaded specifically In addition, we can also click on manual settings Here, the custom allows TrackGPT to perform specific tasks For example We can set it to automatically call daily AI news and GitHub hotline information at 9:00 a.m. and

finally send the compiled content to my flybook in the form of a flybook article.

In this way, I can see the latest AI news and open source hotlines on time every morning. The above is probably the basic play of automation. We've got over 90% of the people here. You can put a 666 on the barrage to show yourself. Finally, let's talk about high-end play. Like hook and hook, work tree, and so

yourself. Finally, let's talk about high-end play. Like hook and hook, work tree, and so on. More of them are used in code development. Click on Settings. Under the code

on. More of them are used in code development. Click on Settings. Under the code selection, you can see that there are hook, link, git, environment and worktree. These

are used more in AI programming. Next, we will develop an AI desktop and deploy it to the stand. Let's take a look at these functions. First, I wrote a development demand article that explained the technical battle in detail. For example, using the full battle framework of Next.js, it is stipulated that the design tone needs to have

the high-end temperament of OpenAI and Apple official website. The main tone uses a deep blue liquid glass style. I also drew this structural trial map in the article and listed the general functional points. We send this required document to AI, let AI develop according to the document, click the plus sign in the lower left corner, use the

Go target mode here. Usually when the document is written in great detail, we can use the target function directly, so that AI can execute the goal once. But if

your task is not clearly described, AI will be easy to run away. So when

using the target mode, the task must be written clearly in the early stage. This

process only needs to wait for a while. But it may take more time and time. In order to save time, we won't wait. Let's take a look at the

time. In order to save time, we won't wait. Let's take a look at the final development effect. The project is generated in this folder. We start it. You can

see that there are calendar, standby, habit card and database functions in the page. The

completion is very high and the UI page is completely in line with my requirements.

It's very beautiful. Usually, using Go orders to develop large projects can generate a demo, and then continuously optimize and change bugs on the demo. Because this AI workstation itself is a small project, there are not many bugs. In the development process, sometimes we have some uncertain ideas. For example, I want to add a login function but I'm

not sure if I want to keep it. At this time, you can click "Send" to enter the new chat. Try planning here. This function is very suitable for trying out the program. But the one that is sent out is still in the same folder. It's generally recommended to use it for plan planning. Another advantage of branching is

folder. It's generally recommended to use it for plan planning. Another advantage of branching is that it does not pollute the top and bottom text. If you encounter topics that you want to discuss in the process of realization, but you don't want this discussion to destroy the top and bottom text in front of you, you can branch out and discuss it separately. By the way, let's talk about the difference between branching to

new chat and walktree. Branching to new chat, the only thing that is divided is dialogue. It will keep the top and bottom text of the previous chat. and then

dialogue. It will keep the top and bottom text of the previous chat. and then

start a new dialogue from the current location. But the two dialogues operate on the same project folder, so it is suitable to change the idea and continue to talk.

But if the two dialogues modify the same file at the same time, they may affect each other. WorkTree is different. It will use Git to create a separate project screen, so that two conversations can be edited and covered in their own work area.

It is more suitable for developing two functions at the same time. In summary, the new chat is a conversation separation. Worktree is a dialogue and project log are separated Here you have to pay attention to Worktree is dependent on Git That is to say, your project must be a Git warehouse first And at least complete a submission

Codex has a starting point to create a new work area. Ordinary folders do not display Worktree. I made the initialization and submission of Git here first, then clicked on

display Worktree. I made the initialization and submission of Git here first, then clicked on the division button. You can see the division creation in the new work tree. After

roughly talking about division, let's see what to do if you want to add an AI chat function to this workstation. AI functions must be included in the API of large model manufacturers such as the official API of DeepSeq We directly told AI to help me add an API key of .env to the DeepSeq AI dialogue function in

the first page I filled it myself because it's more private AI will help us develop it soon. Then we copy the API key. It seems that it can't be filled in directly here. Then I use VS Code to open this project and fill it in again. To be honest, this experience is not very good. It would be perfect if I could fill it in directly here. After fill in, there is a

little error at the moment I sent this wrong screenshot to AI Let him fix it OK, it's fixed Let's test it Ask a question What model are you He replied that it was DeepSeq OK, this means this function is running Just now we modified the .env file Everyone knows that .env files are very important The API key

modified the .env file Everyone knows that .env files are very important The API key inside is absolutely not leaked Otherwise, what others use is your maliciousness Therefore, when AI needs to change this kind of core file, we can actually make a blue line, which will lead to the hook function. Simply put, it is an automatic trigger in

the codex workflow. Run to the designated node, it will automatically execute the script we have prepared in advance. It is particularly suitable for doing those things that are executed every time but are easy to forget. For example, when the project starts, it automatically uploads the top and bottom text. When submitting the request, check if there is any

sensitive information leak. Before executing the order, intercept high-risk operations. Before the mission ends, check if there are any tests and summaries. I will show you a most intuitive security guard. As long as the next operation of AI involves the modification of .env database

guard. As long as the next operation of AI involves the modification of .env database

or core code, I will first make a reminder. After I confirm it, continue to execute. Directly send a prompt to help me write a hook. Before modifying the file

execute. Directly send a prompt to help me write a hook. Before modifying the file or executing the command, I will check it. As long as the next action involves the modification of .env environment variables, database files or core code, give me a warning.

Let me allow it to be executed later. OK, AI has quickly written the hook script for us. We click on Settings to the hook option under the code. You

can see the hook just configured. Let's have a new conversation to test whether it works. Send a prompt: Help me modify the .env file, change the API key to this one. AI is located in the .env file and triggered the hook script. The interface shows a high-risk operation permission warning. Click allow. That's

probably the core function of hook hook Now that the whole project is developed We're going to deploy it to the stop point Send this prompt at the same time This insert key helps me deploy this project to the stop point AI will automatically call relevant skills to check. It says that the current use is the traditional Next.js

service section operation, and the deployment to the site requires a cloud-based structure and a long-term configuration. It will automatically move. We can follow its process. Here, every time it

long-term configuration. It will automatically move. We can follow its process. Here, every time it goes to check the .env file, it will ask us a question. This is the hook and hook that was just set up here. It's working. Wait a minute. The

project finally successfully deployed to the site. Now we can visit the Internet. Check

the page again. Send a message to this DeepSake assistant to see if it works. OK. You can reply normally. No problem. Click on the stop button in the

works. OK. You can reply normally. No problem. Click on the stop button in the upper left corner to see the personal workstation. We can set the full line for only individuals or anyone on the Internet. After the release, you can share this link with your friends. Click on the edit in the workstation and you will return to

the AI dialogue box. Second edit Click on analysis to see the number of views in the page. In the database selection, because it uses its own cloud database, so you can see the data table content. Click on this setting, you can modify the name of the project, URL, and you can also specify the domain name.

The API key environment variable just now is also filled in here. The function of the stop is probably this meaning, it can directly deploy our own developed website online, and save a lot of complicated operation processes. I really recommend everyone to try it.

OK, that's it. This episode of Codex's new version of the Transformation Strategy is all over. We've gone from the basic interface, project folder and task mode, all the way

over. We've gone from the basic interface, project folder and task mode, all the way to office, deep picture, long-term memory, CUI, plug-in, skill, MCP and automation finally made an actual AI workstation. Go, branching, worktree, hook hook and stand-point deployment all ran once.