Codex’s big sale is coming—this official guide will walk you through how to make the most of your budget

Over the past few days,Codex It has received overwhelmingly positive reviews on social media.

Some users have discovered that inviting a friend to join Codex will reset the rate limit.

Even if the invited user isn't a new user or a subscriber, they can still earn a reset opportunity simply by opening Codex via the link and sending a few messages.

In addition to the referral rewards campaign, the official Codex will also be offering significant discounts.

According to foreign media reports citing sources familiar with the matter, OpenAI is consideringsignificantly reduce the fees it charges users...in order to win customers away from its competitor, Anthropic.

The report mentioned that,OpenAI may lower the price of tokens...but discussions about the big price cuts are still ongoing.

After all, Codex is currently OpenAI’s best platform for acquiring new customers.

As per data released by OpenAI, ChatGPT has surpassed 1 billion users, while Codex has just reached 5 million weekly active users,This means that out of every 200 ChatGPT users, only one clicked on Codex in the sidebar.

"Not needing it" is one thing, but more often than not, it’s simply a matter of not knowing how to use it, or what Codex is capable of—specifically, the tasks that ChatGPT struggles with but Codex can handle.

Codex 官方也听到了用户的反馈，一边高调宣传即将并入 ChatGPT，pending我们打开全新大改版的 ChatGPT 应用时，可以选择使用 Codex 还是 ChatGPT 来回答。

Meanwhile, OpenAI has recently updated its official website with more than a dozen real-world workflows in one go. These range from common tasks like deploying web pages and applications, and building Mac or iOS apps from scratch, to large-scale project management, 150-hour research projects, and various day-to-day business tasks—all accompanied by corresponding use cases.

These tutorials are probably the best guides for getting started with Codex quickly; they do a great job of explaining what Codex can do and how to use it.

Computer Use: Let Codex Control Your Computer

Hey Siri, open WeChat and send a message to Mom saying XXXX

Please unlock your iPhone first

Siri can't do it, and neither can Codex at this point.

Codex’s “Computer Use” feature primarily allows AI to interact with computer interfaces just as we do, completing tasks by clicking, viewing, and typing. This feature is well-suited for cross-application tasks, such as collecting notes, updating records, copying details between different locations, and replying to messages.

In the official use cases, the examples range from simply playing a song to switching between different apps.

@Computer, play some music to help me focus.

@Computer, please add my interview notes from Notes to Feishu.

@Computer Please check my WeCom account and set a reminder for me about everything I need to finish before the end of the day.

To use this feature, first locate "Computer Use" in the Codex app and make sure it's turned on. Then, in the chat window, start your command with @Computer, or mention a specific app, such as @Slack or @Messages.

After selecting the Computer Use plugin, describe the task and the desired outcome. When Codex requests access, grant it, and then let it continue running the task in the background.

Here are a few things to keep in mind when using Computer Use: make sure your Mac doesn’t lock while it’s running, enable the “Lock Screen Actions” feature in Codex, and when Codex uses apps on your computer, you can specify the default browser in the custom settings.

Also, avoid using two "Computer Use" task threads to control the same application. Once each thread has finished, you can ask Codex to summarize and optimize the task, or even turn this workflow into a repeatable process.

Give Codex a goal that will keep it going

When we have AI do the work, we usually have to stand by and watch it closely. It takes a small step, then pauses to ask what to do next, so we have to keep a hand on the wheel the whole time.

/goal This is exactly what we want to achieve: give Codex a long-term goal so it can keep moving in that direction on its own, without stopping even after completing a full cycle.

In the official guide, some typical examples are like thisBigger than a single prompt, yet smaller than an entire to-do listthe task of,The goals are clear, they can be verified independently, and it’s clear exactly what constitutes completion..

Project Migration: Whether you're migrating a game to a new tech stack, porting a mobile app to a new platform, or switching the entire codebase to a different framework, you can use `/goal` to have Codex handle the entire migration process.

Prototyping: When building a new app, game, or feature from scratch, you can use the `/goal` command to have Codex generate a polished first draft. You can write a `PLAN.md` file to clearly outline your vision, and Codex will follow your instructions.

Prompt tuning: If you have a test dataset, you can use the `/goal` command to optimize your prompts based on evaluation results. Codex will identify which cases failed, modify the prompts, and rerun the evaluation, iterating until the score improves or until the termination conditions you set are met.

When it comes to writing a goal that can be successfully achieved,First, give it a clear goal and a completion condition; tell it which files, documents, issues, logs, or plans to review first; decide which command and deliverables will serve as proof of progress; have it work in phases, keeping a brief progress log along the way; Throughout the process, we can use `/goal` to check the status at any time; when it’s finished, stuck, or needs to change direction, we can pause, resume, or clear it.

Use GPT Image 2 to create a PowerPoint presentation

The most tedious part of creating a PowerPoint presentation is often the layout. Codex comes with two built-in tools: $$slides uses PptxGenJS to directly read and write .pptx files, while $$imagegen generates accompanying images.

The official prompt provided by OpenAI is:

Using the $$slides and $$imagegen skills, edit this slide as follows:

– If available, please add the logo.png file to the bottom-right corner of each slide.

– On slides X, Y, and Z, align the text to the left and use the image generation feature to create illustrations on the right (styles: abstract, digital art).

– Keep text as text whenever possible, and keep simple charts as native PowerPoint charts.

– Add the following slide: [Describe the new slide here]

– Use existing brand guidelines (colors, fonts, layouts, etc.) in new slides and text.

– Render the updated presentation as slide images, review the output, and fix any layout issues before delivery.

– Run overflow and font substitution checks before delivery, especially when there is a high density of cards.

– When creating a set of related images, save reusable prompts or generation instructions.

Instead of starting from scratch, describe the content and overall style page by page, and put any logos or images in the same folder so they’re easy to access.

We can also have Codex handle weekly, monthly, and quarterly reports by updating templates on a regular basis. We can have it generate a `guidelines.md` file that defines the content, structure, and update process, and then use other tools to pull the relevant data—for example, for quarterly reports to shareholders, we can simply plug in the latest figures and insights.

Alternatively, when editing an existing PowerPoint presentation, you can simply use the dialog box to ask Codex to fix issues such as spacing and text misalignment.

Have Codex create a webpage based on the screenshot

Whether you have a few screenshots, a brief design brief, or some reference images for inspiration, Codex can use them to create a responsive interface while adhering to the existing coding conventions of your project—that is, the existing framework and programming language—without starting from scratch.

When paired with $playwright, Codex can open pages in a real browser, compare them one by one with the screenshots we uploaded based on different screen sizes, and repeatedly adjust them until they look as close as possible.

The prompt is as follows:

Please use the screenshots and annotations I provided as a reference to implement this user interface in the current project.

Requirements:

– Reuse existing design system components and styling.

– Adapt the screenshot-taking utility and components from this repository rather than creating a parallel system.

– Spacing, layout, hierarchy, and responsive behavior should be closely aligned.

– Respect the repository’s routing, state, and data retrieval patterns.

– Make the page responsive on both desktop and mobile devices.

– If any details in the screenshot are unclear, please choose the simplest implementation that still aligns with the overall direction, and briefly explain your assumptions.

Verification:

– Compare the final user interface with the provided screenshots, including its appearance and behavior.

– Use $playwright-interactive to verify that the UI matches the reference, and iterate as needed until it does.

Building a browser-based game from scratch

Game development is probably one of the areas where you can see that Codex isn’t just a coder—they also understand design. A real game requires documented gameplay concepts, a rendering layer, a front-end shell, a back-end state system, and art assets, and it also demands constant tweaking of the visuals and gameplay feel.

Before you start building the framework, have them write a PLAN.md file that breaks the game down into specific components: player objectives, core loop, controls and input, victory and defeat conditions, difficulty and progression, visual direction, tech stack and deployment assumptions, and the sequence of milestones.

Create another AGENTS.md file. Follow the official tutorial; you can use the example below as a reference.

Game Title

<游戏类型>

Technology Stack:

– Frontend: Next.js (deployed on Vercel)

– Rendering using

– Backend: Fastify + WebSocket (deployed on )

– Postgres for the database; Redis for caching and pub/sub

– Generative AI features from OpenAI

Agreement:

– After completing each feature, run the `build / test` command to verify it

– When developing new features, follow the PLAN.md file

– Record your thought process and decisions in the .logs file so you can refer back to them during iterations

– Use Playwright to test the visual layout; if it doesn’t look right, make adjustments

– Use ImageGen to generate images; after each batch, save the prompt to the .prompts folder so you can easily generate more of the same type later.

– Use Context7 MCP to pull the documentation

Install all the tools mentioned in AGENTS.md: $$imagegen for generating art assets, $$playwright for testing the game in a real browser, $openai-docs to fetch the latest OpenAI API documentation, and if needed, add Context7 MCP to fetch the rendering framework documentation.

Next, Codex will proceed as planned and generate the first version. If there are a lot of images to generate, this version might take several hours to run, and Tokens will start burning like crazy. However, thanks to Playwright’s capabilities, Codex can test and verify the game’s performance in the browser on its own, with virtually no intervention required from us. The more detailed the plan, the more polished the first version will be.

We had Codex write a Plan.md file for a game on its own; we entered a prompt, and it generated a mini-game that was almost ready to launch.

Use $playwright-interactive, $imagegen, and $openai-docs to plan and build a browser game in this repository. Implement PLAN.md, and log your work in the `.logs/` directory.

In addition to small web games, using the iOS app-building plugin provided by Codex, we can view and test iOS apps right within Codex with just a single command.

Let AI Conduct Research on Its Own

Codex isn’t just good at writing code; it can also serve as a long-term research assistant in scientific research. Users provide direction and guidance, and Codex implements, validates, evaluates, and iterates repeatedly.

One example involves modifying the model architecture. Suppose you have a hypothesis regarding protein folding: “If the model were to represent more higher-order geometric structures, would it learn better?” However, this kind of idea can’t be fully explored in a single attempt; it requires repeated experimentation.

With Codex’s Goal Mode, give it three things: a clearly defined scientific direction, a baseline model that works, and a set of automated evaluation metrics. It will then work toward that goal, continuously improving its performance through implementation, testing, logging experiments, troubleshooting, and iterative refinement.

In the official example, Codex ran continuously for over 150 hours, producing an experimental architecture called SimplexFold.

Another challenge is ranking drug targets. The difficulty with tasks like this lies in the fact that the evidence is scattered across a dozen or so databases, with genetic, clinical, literature, and expression data each kept in separate repositories.

Using the Life Science Research plugin, Codex can simultaneously retrieve evidence from various databases, assign a score of 1 to 5 to each piece of evidence, and finally compile the results into a single scoring table and a ranking, complete with visualizations such as heat maps.

The OpenAI website lists many more use cases; here, we’ve only highlighted a few of the most popular ones. If you’re interested, you can visit the OpenAI Developer website.developers.openai.com/codex/use-cases, try out different scenarios.

This article is from WeChat“APPSO”, Author: APPSO—Discovering Tomorrow’s Products