All weekend, I was almost asleep, all the other time in Vibe Coding。
You know, you're writing a story while you're working on it, you're cooking on it while you're watching a movie。
Really, the Macbook on the lower left is the computer I used for Vibe Coding, and the big screen is the Windows mainframe I used to watch movies。
It's too simple to use Mac for CodingCodexIt's an application。
Mrs. Codex+GPT-5.3 Codex is fine。
I SENT AN X ON SATURDAY。
There's a lot going on down there。
really, fast and awesome, 5.2-codex used to be bad, not because it wasn't strong enough, but because it was too slow to go out and hang the coding and go home for two, three hours。
But this time, 5.3-codex is twice as fast, is high, has a higher capacity than Opus 4.6, my own body feeling and experience, and really, it's much better than Claude alone。
To be honest, Vibe Coding was interested in the past, but because Claude's cover was discouraged, because OpenCode's various BUGs were discouraged, because Ide did not understand the people who were discouraged。
Today, I can say on my chest that the Codex+GPT-5.3-codex of OpenAI is your best entry, entry and graduation product。
You have to believe me, as foolish as I am, that I can feel the pleasure of getting into the flow of heart, using it one weekend to meet the development needs of four or five of me that I was unable to achieve by myself。
Someone like me can do it, then you can do it。
I would recommend that you start at this moment, no matter how white you are, you can start using AI, using codex to meet your own needs。
And, the most important thing is, Codex has the app! He has visual interfaces
As a user experience designer who's been doing it for almost 10 years, seriously, as a non-professional programming user, I can't really accept it for 2030, and I'm going to use the command line interface。
Really, you know the first time I used the command line, and when OpenCode's TuI, I didn't know how to open a directory, or even have an extremely unusual thing, which is to copy paste。
We both know that Ctrl+C is a copy, right? That's what I thought。
And then in the command line, I want to copy something, and I'm used to it directly, and then press Ctrl+C, and this shortcut is forced to close。
And Claude's XX, not only with the hysteria, but also with the feeling that the money really fell out of the eye, and this weekend they gave Opus 4.6 another Fast mode, which is 2.5 times faster。
Everyone thinks it's okay, so when I look at the file, it's five times more expensive. Token, 1-M, consumes 150。
Turn the other side of Otmann。
I love research, but coach, I really don't want to suffer。
There's a saying that this wave I'm standing on OpenAI, I love OpenAI。
So, today's article, I sincerely recommend Codex to you, and I hope that you can really start your vibe journey in this era。
it's just in time for the spring festival, and it's better to be free at home, to create it with coding, to show it in front of relatives。
What's Codex
Let's do a little Cope first, because many friends may not have heard of it at all。
I'm sure many friends know Claude Code and Clawdbot, both of which can be understood as an Agent application, with a lot of engineering capability on the upper floor。
And Codex, it's a Claude code product from OpenAI against the Anthropic family。
A programming agent, but by now it's almost equal to Universal Agent, because it's 30 years of computerization, and almost everything is built on code, and the more you're programmed, the closer you are to Universal Agent, which is why the logic of software has been so noisyly changed around the U.S. stock lately. An excel coworker plugin in Claude can scare Wall Street half to death。
both codex and claude code are top-level programming applications, which in themselves are to be used in combination with a matching model, and both of them sent their own new models, GPT-5.3-codex and Claude Opus 4.6 last Friday。
GPT-5.3-codex is a purely programming-specific model, so it is not working well in terms of creation, fact-checking, world knowledge, etc., so OpenAI did not put him up for use in ChatGPT。
ChatGPT is only GPT 5.2。
Only Codex is now connected to GPT-5.3-codex。
If you're a Plus or Pro member, you can download the codex and then use it。
If you're a free-of-charge member or an eight-dollar Go, Codex can download or use it, but without GPT-5.3-codex, you can use GPT-5.2-codex, which requires attention。
As for how to hit Codex, that would be easier。
Into OpenAI's Codex network:
https://chatgpt.com/codex
See this big download app? Just click。
A Mac installation package is downloaded and installed and log-in is finished。
Not too simple。
As for the Windows application, it is not on line at this time, but only the command line version, but the application should be fast, it is expected to be in the week。
II. How to use it
And when you download it, and you log on to Codex's front page, you should see that this is the interface。
In the logic of programming, there's actually a very central thing, which is the sidebar on the left, a one-by-one folder, something called Threads, which is a thread。
In the whole Codex logic, the left-hand column actually has two layers。
The first level is a folder, a working area。
You can understand it as a catalogue of projects, or a subject box, like my own AI hotspots, data grab robots, sandbox, and so on, it puts your file in order。
The second floor is Thread。
You click a folder, you see a conversation in it, and those conversations are Tharead. Each Thread is a separate line of task。
So the whole relationship is this。
There can be many Threads in a folder, each of which is in the same working area, a collaborative process around a clear goal。
Take the best-tempered metaphor。
Folders are like a project group。
Thread is like a specific topic in this group。
You're talking about demand in one subject, Codex is changing files in the same context, running orders, making records, and you're going to change the subject to another task line。
It's really nice for White, because it's tearing apart two things that are easy to mix。
The folders are for code and information, and Thread for ideas and processes。
You're not going to experience that very classic crash。
Let it write a page in the morning, let it count Excel in the afternoon, let it change at night, end up with a pan of porridge, all of it in one piece, the context was extremely polluted, AI started making up a mess, and it couldn't find the file。
In Codex, you just follow a simple, ridiculous rule。
In the same folder, you can do the same big thing, and in the same Thread there is only one specific thing to do, which is best。
They share documents and resources with each other in the same project catalogue。
However, their dialogue and objectives are non-polluted and can be intermittent at any time。
So, I very much suggest that we start with a sort of classification。
i'm a very simple habit, for example, of building a folder called dev on my computer。
There's some of this stuff that Learning put in my learning materials, nottes is some of my own articles and notes, and Projects is the real project that I actually do, and Sandbox is the sandbox, and stuff that I don't know how to sort out can be thrown in here. Tools is my own generic scripts, reusable components, little tools, etc。
There are, for example, a few real missions in Project, Ai hotspots and flying book robots, and the flying book robots have several different functional fly book robots。
For example, just finished on the weekend, the flying-book robot that crawls down my public number data on time into my multi-dimensional forms。
Don't look at me talking a lot, but he's really important! A good classification is the beginning of a happy follow-up. Don't ever talk to ChatBot。
When you've built it locally, you can add a folder to your project folder。
For example, I'd like to develop my AI hotspot site, so A can add this project folder and open a Thread for dialogue。
At this point, if you want to say anything, you can just send a message。
But I know you're in a hurry, but you're in a hurry and there's some configurations and functions, and you can start playing after I'm done。
III. Functions and configuration
After all, Codex is a great product, and there are some features and configurations, and I'll finish by introducing you。
And that's what the graphical interface brings, and it's very friendly for our whites。
The first is a time-bound mission。
It probably means that Codex would do something specific on a particular date。
i had one myself, for example, because several of my projects were on the cloud server on my volcano engine, and i didn't really know the server at all, so i just gave it to cordex, which deployed and operated everything, and it built itself an automated, 9 a.m.-a-day inspection of my server to see if it was wrong。
If you make a mistake, you'll sort it out yourself, and then you'll sum up the reasons and send it to me through a flying book robot, so that I can have complete hosting。
the second one that i think is awesome is skills。
i don't think i need to explain how important and awesome the skills are. the old friends who read my article must know that i wrote a lot of skills lessons and shared them。
and for the first time, skills had his own visual, graphic interface。
you don't have to be like claude code or opencode, you don't even know what you've got skills and where you've been。
All you need is this interface, and you can manage it easily。
Meanwhile, Codex brought his own skill Creator, which means that building a skills on the codex is almost the same as a button, and you just have to click on New Skill at the top right corner。
then use your mouth to tell him what kind of skill you want to build。
It's so convenient that Claude Code and OpenCode are so stupid that they support Skill, but the Skill Creator who built the Skill needs to go and install it myself, which is stupid。
OpenAI, still understand users。
In addition to these two visible functions, there are a number of small configurations that can be modified。
Enters the settings interface。
In General, turn the switch on to keep the computer on, replace-up behavior with the steper, so that you can send a message to Codex at any time during development。
And then in the Personalization (personalization) section, you can fill in the global rule, which is actually AGENT.md in the order line, except that OpenAI was brought out alone。
This is the global rule, and I'd like to recommend one that I'm used to
# Global rules for codex
##
- Prefer small, highly qualified.
– Before edifying, independent of the file (s) to change and state the plan in 3-6 bullets.
- Ever invent APIs, configs, or file paths.
- Keep changing with exciting style and analytical.
#1TP5Safety and security
- No, I'm sorry.
– If a task involves security, ask me to provide them with information.
- Do not add analtics, please, or network calls unless I ask.
## CodeQuality bar
– Add or update tests for behavior changes when the project has tests.
-Prefer type safe and elaborateer handing.
– Add comments only when the int is non-obvious.
## Build and run etiquette
– If you need to run companies, develop the exact company and why.
– When you make change that may break built, run the most effective check first.
##
– For code changes: include a short response + list of files changed.
– For debugging: include hypotheses, examples run, and the mini-fix.
## My references
– I like concretifications, concrete steps, and copy-pastable committees.
– Different language for examples: Chinese.
OKAY, BY THIS POINT, SET THIS PIECE UP。
Then, on the front page of the dialogue box, change the permission to Full Access to give Codex the highest access to your computer, so that you don't have to ask for permission every time。
Every time you go through a piece of Prompt, you open up。
Then in the dialogue box, you can output/key, and you can hang some special features。
There are two important new players。
First is Plan Mode。
It's a simple function. He only plans, doesn't write a code line, and then he picks out a small icon。
From 0 to 1 for each large project (to download a video, to transfer it), I recommend you, using the Plan model, to plan your needs in detail, to create a standard file and realize the plan, and then to develop it。
For example, I'm going to build a skill that manages the source of my AI hotspots, because each one of them is different, and I choose to manage it directly by using a skill to modify my code library。
at this point, you can plan for the skills and then start developing the skills。
You'll get the plan in a while。
at the same time, the codex will ask you whether this plan is being implemented, yes and no, and generally you choose。
Then his little Plan icon disappears and officially enters development time。
You just wait and finish。
the second is status, and it's very useful to see the amount of your current cycle and the amount remaining。
Oh, there's this piece of model reasoning that needs to be singled out。
GPT-5.3-codex is, as before, four levels of reasoning。
you can understand that the higher the slot, the more thought token and self-censorship the model before giving the final answer, the more steady, but slower and more expensive。
I recommend you to use High on a daily basis, and once you think you're going to go beyond normal, you really have to do some hard work, you can use Extra High, which is basically 5.2 xheigh。
Such great work, the success rate and the time cost are the most valuable, and some of the additional reasoning is consumed and ignored。
IV. Start Vibe Coding
finally, we can finally start coding。
there's nothing to say about this piece of coding, so just tell me what you need。
There's a microphone in Codex's lower right corner, and I'm always opening it now and then talking。
My personal development habits are like this:
Then go on to codex, open the Plan mode, use my mouth to describe my needs, and when it comes to front-end pages or servers, emphasis is placed on the use of front-end design skill, i.e. Frontend Design, for design, and my own server management skill for intervention management。
Once the plan has been produced, I will develop it, and when it is finished, I will first see if the front-end effect is feasible, and if it is, then I will make subsequent changes with my mouth in the codex, and if the effect is a piece of shit, I will open my claude code directly to the same project warehouse, choose Claude Opus 4.6, and let him reset the front-end。
This is true because sometimes Codex's front-end capacity is really very general, even as Otmann himself says。
If you don't have Claude opus 4.6, it's no big problem with claude code+K2.5+Frontend Design skill。
once the front end has been reset, we can go back to the codex and continue to be happy。
Oh, yeah, there's a little trick。
In the codex, it can be developed in parallel with multiple Threads。
i hope you're happy to coding。
In my last article, I said
I always thought Vibe Coding might be more valuable to non-programmers than programmers。
BECAUSE PROGRAMMERS WOULD HAVE WRITTEN CODES, AI WOULD HAVE WORKED FOR THEM, BUT FOR THOSE OF US WHO COULDN'T WRITE CODES, AI WOULD HAVE JUST LEVELED OUT AN OTHERWISE UNATTAINABLE MAGMA。
In the future, it's going to be a basic skill to write with an AI。
It's a necessity。
It is to be hoped that everyone will be able to use their own ideas。
Have fun。