Codex App Practical Guide: Earning While Playing – The Hidden Tips No One Talks About

In my last post, I covered Codex App from zero to one, focusing on getting everyone familiar with the Codex interface and basic configuration.

逸尘@gengdaJ·May 6ArticleCodex App Complete Beginner Tutorial: Breaking Down Every Detail of This Super AppLately I’ve noticed that when many people first encounter Codex App, their reaction isn’t “Wow, this is amazing” but rather: “Where do I even start with this thing? Do I need to configure a bunch of stuff?”…444772.2K955K

Now that the foundation is laid, it’s time to use Codex App to do things that truly boost efficiency and reduce costs, fully integrating it into your workflow.

Today, I’ll walk you through practical scenarios to experience those lesser-known tips and tricks. From basic to slightly advanced, from features to real-world applications, I’ll help you master Codex in one go! 🔥

• Essential for working folks: Use Codex to handle Word, PDF, PPT (image-based vs HTML-based), and Sheets – the big four office tools.
• DIY cool 3D videos: Use Codex + HyperFrames / Remotion (three.js) to create “Top 10 Chinese Animated Films.”
• Healthy routine: Use Codex to build a beautifully designed website analyzing the benefits of early sleep and early wake, deploy it on Vercel, then use Codex to control OpenRouter for comparative experiments – test the new Ring model.
• Must-have for content creators: Use Codex to create long image-and-text tutorials – let Codex control itself, run multiple conversations in parallel, take screenshots, blur and go.
• E-commerce product listing automation: Use Codex + Playwright to control the browser, painlessly batch-upload products on Pinduoduo.

Learning while playing, and then earning while playing – that’s my goal, and I share it with you. Enough talk, let’s dive in!

1. Essential for Working Folks: Using Codex for Word, PDF, PPT, and Sheets

In the past, to create an event plan, project report, course material, consulting report, or employee training document, you’d need months of practice with these tools.

But now, with a single prompt in Codex, you can generate beautifully formatted Word, PDF, PPT, and Sheets with accurate data.

To showcase Codex’s ability to produce the four office file types, I created a demo:

In the current conversation, please create an office suite demo with the theme “Codex App Content Production Workflow.” Requirements:

1. Generate real files for Word, PDF, PPTX, and XLSX – four types.
2. Each file must have actual content, not empty files.
3. After generation, tell me the real file path for each.
4. Preferably generate a preview or file list so I can screenshot it for the tutorial.
5. Finally, provide an acceptance checklist confirming all four files exist, can be opened, and are not placeholder content.

Verification results:

1. Word document verification screenshot:

2. PDF file verification screenshot:

3. PPT presentation verification screenshot:

4. Sheets spreadsheet verification screenshot:

Even without giving a custom skill or optimized prompt, achieving this result is already excellent!

As we all know, the hottest PPT creation method on X right now is definitely the revolutionary HTML-based PPT. Related trending skills are popping up everywhere. Another creative approach is generating image-based PPTs.

Below, I’ll use the “advanced” PPT workflow to give you some inspiration:

Route 1: Image-based PPT

Image-based PPT means: First, have Codex generate a series of 16:9 visual images, then lay them out in a PPT.

Each slide is essentially one complete image.

The effect is pretty good, since it’s powered by GPT2-image underneath.

Of course, PPTs created this way aren’t completely uneditable – there are still ways. If you’re interested, feel free to DM me for the document.

Route 2: HTML-based PPT

HTML-based PPT is a very popular route recently.

Simply put: instead of using traditional PPT software, you write slides directly with HTML, CSS, and JS.

Its advantages:

• Preview directly in the browser.
• Add animations.
• Add interactions.
• Control layout with code.
• Easy to screenshot, export, and screen-record.
• Perfectly compatible with code-execution tools like Codex.

There are many popular open-source skills on X. Just search with Grok, and Codex can install and use them.

2. DIY Cool 3D Videos: Codex + HyperFrames / Remotion – “Top 10 Chinese Animated Films”

The previous section was for working folks. Now let’s play something cool – a balance of work and play.

I previously used Codex + HyperFrames / Remotion (three.js) to create a 3D video of “Top 10 Chinese Animated Films”:

逸尘@gengdaJ·May 7Codex+Hyperframes+Remotion Generate the “Top 10 Chinese Animated Films” Ranking High-Energy Mashup in One Go! It’s fire!

1. Minimax voiceover narration
2. yt-dlp to download part of the video
3. Hyperframes+Remotion to generate opening animation and edit
4. Local tool to normalize volume

All done by Codex in one attempt!!!Quote逸尘@gengdaJ·May 6ArticleCodex App Complete Beginner Tutorial: Breaking Down Every Detail of This Super AppLately I’ve noticed that when many people first encounter Codex App, their reaction isn’t “Wow, this is amazing” but rather: “Where do I even start with this thing? Do I need to configure a bunch of stuff?”…30208883136K

Core idea:

• Codex handles writing projects, modifying code, running rendering, and debugging errors.
• three.js handles 3D visuals.
• Minimax handles voiceover narration.
• yt-dlp downloads part of the video.
• Remotion / HyperFrames converts web animations into video.
• Local tool normalizes volume.
• Finally, output MP4.

Detailed steps:

1. First, clearly tell Codex your video requirement – “I need a video for ‘Top 10 Chinese Animated Films’, 16:9… (more description refer to template below). Go find source footage online and download it.”
2. Align requirements with Codex – “If anything is unclear or requires my judgment, please stop and ask me questions for alignment.”
3. Enable “Ultra” and “Full access permissions,” let Codex @browser and @computer use do the work. @chrome also works (though I’ve always timed out and failed).
4. Get the output, then tell Codex what you don’t like and have it make changes.

When I first wanted to create this video, my thinking was simple: I wanted to verify that Codex’s chain – “auto-write script → search for materials → auto-download → auto-identify highlights → auto-narrate → auto-edit” – could work.

Then, shifting to commercialization: Call APIs like GPT2-image, Seedance 2.0, ElevenLabs, and automate 90% of the above workflow, applying it to real scenarios like AI-driven traffic monetization, AI e-commerce promotion, AI front-end customer acquisition, and AI advertising.

As for how to write scripts, generate images, and create videos – that’s a major discipline requiring countless trials and errors. It can’t be learned overnight. Moreover, the skills required for each vertical domain are completely different.

But there’s a universal template for quickly generating a video and getting started:

Goal: Audience: Target platform: unspecified / YouTube / Shorts / Reels / TikTok / Website / Conference venue Duration: Aspect ratio / resolution / FPS: Language: Number of scenes and duration per scene: Subject / Character: Actions: Scene / Background: Camera / Composition: Lighting / Color / Style: Subtitles / On-screen text: Audio: Narration / Language / Voice / Speed / Music / SFX Transitions: Keyframes / Reference images / Character consistency assets: Brand / Legal / Accessibility: Data variables / Placeholders: Output requirements: First provide scene list, then final prompt, then render metadata JSON.

For pure AI video, my previous post also shared a summary of basic film knowledge:

逸尘@gengdaJ·Apr 26https://bcn5ot9wwnew.feishu.cn/wiki/KttewP2WqiWF0Bk8P68cAjTUnFh?from=from_copylink… A “Basic Handbook for AI Directors” for you! Covers: shot scale, camera angles, composition, lighting, style, special shots, camera movement. After learning, you’ll instantly become an AI master director!Quote逸尘@gengdaJ·Apr 26GPT2-image+Seedance2.0 is definitely game-changing for the advertising industry! Feel the quality and movement of these four perfume videos! Can you believe I have zero art or film background?!

All videos, Made By Rita:https://gamsgo.com/ai-video/partner/WsMzs…416263554K

3. Early to Bed, Early to Rise: One Prompt to Build a Website and Deploy on Vercel

Many people playing with AI now share a common problem: irregular sleep schedules.

The reason might be… AI is just too fun – I originally wanted to tweak one prompt, then casually scrolled X, somehow called Codex to run a task, and then that task wouldn’t stop… That was me in the past.

I know using AI is great, but your body can’t take it – health is the foundation of everything. So while writing this “Codex coding tutorial,” I had Codex build a website with the theme “Early to bed, early to rise, makes a person healthy, wealthy, and wise.”

Having Codex code a small website or app is really simple:

Please generate a new website for me, with the theme “Early to bed, early to rise.” Requirements:

1. The page should be beautifully formatted.
2. It can have 3D animation effects.
3. Analyze the benefits of early sleep and early wake.
4. Provide actionable steps for ordinary people.
5. Make it a real, runnable website.
6. Finally, start a local preview and deploy to Vercel.

Regarding online deployment (others can access your website via a link), you may need to configure the Vercel plugin (search “Vercel” in the plugins section):

Codex didn’t just produce a static page (already deployed at: https://codex-1-html-css-js-2.vercel.app/, feel free to visit). It also created:

• A 3D circadian rhythm visual scene.
• Explanation of benefits of early bed and early rise.
• Evening shutdown routine.
• Morning startup routine.
• 7-day adjustment plan.
• Self-assessment quiz.
• Common misconceptions.
• References and disclaimers.

Of course, this is a basic website. If you want a full-stack website with front-end and back-end, the difficulty increases. But Codex can handle it – remember to choose “Plan Mode.” Having a plan first, then writing code, yields much better results.

Besides having Codex write its own projects, you can also leverage the official web-based AI to generate code – I just open the corresponding web AI, tell it my project requirements, and ask it to return the code. This trick works great when testing new models.

OpenRouter often has new models. Many people are lazy and just wait for AI bloggers to test them, but then they never truly understand a model’s real capability.

Testing one by one is time-consuming. Here’s a little trick – directly let Codex test models on OpenRouter.

Recently, a new model appeared on OpenRouter: Ring-2.6-1T (free), Ant Group’s first reasoning large model.

I wanted to connect Ring-2.6-1T to Claude Code, then let Codex operate Claude Code to see the model’s real-world performance in agent programming scenarios. But I hit a limitation: Codex currently cannot directly control the Ghostty / Claude Code terminal interface already open on my machine. So that chain was difficult to test.

So I took a more direct approach: let Codex open the OpenRouter website, select Ring-2.6-1T (free) in the Chat Playground, then feed it the very ordinary “early sleep early rise” website prompt combined with aesthetic principles from my frontend-design skill, and ask it to directly generate a single-file HTML website.

I also deployed the version it generated. You can view it directly:

https://ring-sleep-site.vercel.app/

Honestly, the result is pretty good – clean and beautiful. Successfully freeloaded tokens!

So, if you want an efficient way to test a new model’s capabilities, I suggest trying the Ring model for free via Codex + OpenRouter. Domestic large models are now not just about parameters and leaderboards – they’re quite capable for real work.

Moreover, this idea can be extended. Codex only uses GPT 5.5, but the web version of GPT can use GPT 5.5 Pro, which is said to write better code than Codex. So extending this idea to GPT 5.5 Pro is very powerful! Don’t let this secret out – just use it ourselves…

4. Must-Have for Content Creators: Using Codex to Create Long Image-and-Text Tutorials

This section is a very common scenario for me, and I believe it will help many friends who are afraid of writing long articles.

What’s the most painful part for content creators? It’s not the inability to write, but the “search for materials → find topics → structure → write → add images/videos” process taking too much time. Often, writing one article takes hours…

So, I simply let Codex do the writing. My previous article that went viral with 920K reads – “Codex App Complete Beginner Tutorial: Breaking Down Every Detail of This Super App” – was actually 80% done by Codex!

Including the article, screenshots, blurring, and cover image. My role was only the 20% correction and supplementation.

But why do many people write junk with AI… while I can produce high-readership articles with Codex? This involves “prompt accuracy” and “style transfer learning.”

Writing is not about giving AI a direction and letting it figure things out. You need to tell it your thought process for writing that article, and let it accelerate the process. The core and content of this article are almost identical to what’s in your mind; AI just automates the tedious steps.

Therefore, before a good AI-written article, there is a person with a clear picture and ideas in their head.

Take my previous article as an example: I knew I needed to introduce Codex’s main interface and settings. So my instruction to Codex was:

“You will operate the Codex APP itself. Analyze each button on the main interface one by one, then open Codex settings and follow the left navigation bar order, page by page, carefully introduce each setting. Be sure to blur any private areas. @computer use”

Then, I collected my past writing experience into a Skill, so Codex understands my writing habits and language style, producing articles that better match my own level. For instance, I created a “Yichen tutorial writing” Skill (extracted from past long articles) and let Codex call this skill to help me write.

Here’s another practical demo – you can see the effect:

An even more powerful operation: let Codex open several conversations simultaneously to run tasks, monitoring itself. For example, the first draft of this very article was completed by Codex opening five windows, greatly improving efficiency!

5. E-commerce Product Listing Automation: Codex + Playwright for Batch Upload on Pinduoduo

Finally, let’s talk about a business scenario that truly reduces costs. I previously had a live chat with an e-commerce store owner. He mainly does business on Pinduoduo, and his store is expanding rapidly.

Their pain points are very typical:

• Huge number of product links.
• Multiple SKUs under each link.
• Thousands of rows in a spreadsheet.
• Hundreds of currently active links.
• When opening a new store, they need to create hundreds of product links from scratch.
• Each SKU requires filling in stock, group purchase price, specification code, product code, reference price, etc.
• Daily maintenance of prices and code changes.
• Manually entering data from spreadsheets into the backend is slow and error-prone.

This is a perfect scenario for automation because it’s not creative judgment.

It essentially is:

Stably inputting structured data from a spreadsheet into a web backend.

First Approach: Computer Use

My initial suggestion was to use Codex’s Computer Use.

That means letting Codex directly look at the screen, control the computer, open the spreadsheet, recognize backend fields, and fill in the content.

Since Codex can operate the computer (I’ve smoothly run database tasks and long X article tasks with it before), I recommended this first.

As long as you clearly describe the spreadsheet path, backend page, and field mappings, Codex can operate like a person looking at the screen.

However, Codex refused to access the Pinduoduo backend, so we had to switch to the second approach – using Codex to call Playwright!

Second Approach: Playwright

What is Playwright? You can think of it as “making code control the browser.”

It doesn’t rely on “seeing” the screen; it directly identifies buttons, input fields, and dropdowns on the web page.

This is actually more suitable for batch tasks. While Computer Use is like a person sitting in front of a computer looking at the screen and operating, Playwright is like an automated script that directly finds web elements without stealing your mouse and screen.

Then this approach worked:

flowchart LR A[“Product base table”] –> B[“Organize field mapping”] B –> C[“Playwright opens backend”] C –> D[“Fill product basic info”] D –> E[“Fill SKU prices and stock”] E –> F[“Upload images and videos”] F –> G[“Screenshot and pause for manual confirmation”] G –> H[“Batch run after confirmation”] H –> I[“Output success/failure log”]

Codex’s Role Here

In this project, Codex didn’t just call Playwright; it handled the entire pipeline:

• Read the original Excel.
• Analyze which columns are useful.
• Reorganize into a standard listing table.
• Let you confirm field mappings.
• Write the Playwright automation.
• Open the Pinduoduo backend using an already logged-in browser.
• Run one product first.
• Screenshot and pause for manual confirmation.
• Record successes, failures, and error screenshots.
• Once stable, run in batch.
• Finally, save the process as a Skill.

Turning successful experience into a Skill upgrades from “help me fill in this one table” to “a reusable automation process for the company,” greatly reducing future trial-and-error costs.

6. Common Pitfalls and Solutions

After running through these five scenarios, I’ve encountered more pitfalls than success stories 😅. Below are the most representative ones, which may save you a lot of time.

Pitfall 1: Computer Use Is Not a Silver Bullet

Symptom: Asking Codex to use Computer Use to operate the Pinduoduo backend results in a direct access denial.

Root cause: Some websites detect automated operations (browser fingerprint, WebDriver markers, headless mode, etc.). Platforms like Pinduoduo have particularly strict anti-bot mechanisms. Computer Use essentially “looks at the screen + simulates mouse/keyboard,” which still differs from genuine human behavior.

Solutions:

• Prefer Playwright: It doesn’t simulate mouse/keyboard; it uses Chrome DevTools Protocol to control the browser kernel directly. Combined with stealth configuration, it appears more human-like.
• Reuse an already logged-in browser: Let Playwright connect to a Chrome instance you’ve manually logged into (--remote-debugging-port), bypassing login checks.
• Division of labor: Computer Use is suitable for open websites (Google search, document editing, posting on X). Playwright is suitable for backend systems with anti-bot protection.

Pitfall 2: Codex Cannot Control Terminals Already Open on Your Machine

Symptom: Trying to have Codex operate a local Ghostty / Claude Code terminal window – it can only see the screen, not take over input.

Root cause: Codex’s Computer Use can recognize screen content but cannot send commands to external terminal processes. Also, terminal application interface elements (cursor position, input focus) are a black box for automation.

Workarounds:

• If your goal is to “use the capability of a web-based AI” (like ChatGPT Pro, Ring), let Codex directly open the web version.
• If your goal is “have Codex call Claude Code,” it’s currently not feasible – wait for Codex to expand its whitelist. Consider using Codex directly for tasks Claude Code would do.
• Execute CLI commands directly in Codex’s built-in terminal – that’s possible, as Codex has full control over its own terminal.

Pitfall 3: Giving a Direction Without a Blueprint → Garbage Output

Symptom: Telling Codex “Write an article about XX” produces junk.

Root cause: Codex doesn’t know your knowledge structure, aesthetic preferences, or target audience’s cognitive level. You give it a direction, it free-wheels; you give it a blueprint, it accelerates execution.

Solutions:

• Before writing, create your own framework: core idea → article structure → key points for each section (one sentence each).
• Feed the framework as a prompt to Codex, asking it to “accelerate the process,” not “create from scratch.”
• Save a writing style Skill (extracted from your past high-quality articles) so Codex mimics your tone and formatting habits.

7. My Understanding of Codex After These Five Scenarios

If you only use Codex as a chat assistant for casual conversations, that’s a huge waste…

You can use it to produce text, images, and videos – write image-and-text tutorials (for content creation), make beautiful Hyperframes videos (content creation), AI video promotion, accept ad orders.

You can use it to parse WeChat chat data (https://github.com/mcncarl/yichen-skills/tree/main/wechat-local-vault) – empower private domain, improve scripts, discover business opportunities.

You can use it to automate browser control – e-commerce AI efficiency boost, multi-platform content syncing.

You can use it for AI showdowns – open web pages and simultaneously call the strongest models from Gemini, ChatGPT, Claude, Grok, etc., for in-depth research or thinking.

You can use it to automatically test new models on OpenRouter and see the capability boundaries of each mysterious model.

You can use it to call and supervise other AI models doing work – this potential is limitless. The only regret is that Codex cannot supervise other terminals (i.e., cannot supervise Claude Code). Hopefully, the whitelist will be expanded in the future.

8. My Suggested Learning Path for Beginners

If you’re just starting with Codex App, don’t jump straight into the most complex e-commerce automation.

I recommend this order:

• First, generate a few images. You can learn prompts from 苍何老师’s prompt website (https://gpt-image2.canghe.ai/).
• Then generate Word, PDF, PPT, XLSX.
• Then build a small HTML webpage.
• Then start a local preview.
• Then deploy to Vercel.
• Then create image-and-text tutorials.
• Then try HTML PPT.
• Then try video animation.
• Finally, touch business automation.

This order is smoother because you’ll gradually build trust with Codex – first see it generate materials and websites, see it go live, see it control the browser.

Finally, you’ll truly understand: Oh, Codex App is not just a chat box – it’s an all-powerful super assistant!

9. Pricing and Usage Notes

Regarding Codex usage: The free version gives you two conversations, then you’re out. So to get started, you’ll generally need at least a GPT Plus subscription. For heavy usage like mine, I opened a $100 plan, which I found lasts about a week in practice.

If you need GPT Plus or Pro top-ups, feel free to contact me (yichen10801 or yichen365ai). I’ll save you recharge time while offering you a near-cost price – just to make friends.

If you encounter the following situation during conversations with Codex:

The solution is:

First, don’t only configure HTTP_PROXY, HTTPS_PROXY, ALL_PROXY. You also need to add WSS_PROXY and WS_PROXY for WebSocket, because new conversations may use wss:// connections. Second, write these proxy variables into ~/.codex/config.toml, so Codex’s child processes can use the proxy. Third, use launchctl setenv to write the same proxies into the macOS GUI app launch environment, so Codex Desktop opened from Dock/Finder can also access the proxy. Fourth, after configuration, you must completely quit Codex and reopen it – not just close the window, because old processes won’t automatically inherit the new proxy environment. Core idea: simultaneously cover HTTP proxy, WebSocket proxy, and Codex Desktop’s own startup environment.

Finally, I hope you enjoy using Codex! If you have complex business scenarios that need optimization, feel free to reach out.