OpenAI has published more than a dozen real-world Codex workflow case studies on its website, covering everything from deploying web pages and building apps to managing large projects and 150-hour research tasks, each with a corresponding method. Codex is about to get a substantial price cut, making now the best time to learn it and get hands-on.

Official case collection: developers.openai.com/codex/use-cases

Before You Start

• You need a ChatGPT Pro or Team subscription (Codex opens from the sidebar)
• Know your way around basic terminal operations
• Codex can consume far more tokens than an ordinary ChatGPT conversation; start with small tasks

Computer Use: Letting Codex Operate Your Computer

Computer Use lets the AI operate your computer's interface the way you do — clicking, viewing, and typing to complete tasks across apps.

How to Use It

1. In the Codex app, find Computer Use and confirm it is enabled
2. Start your message with @Computer, or mention a specific app such as @Slack or @Messages
3. Describe the task and the expected result
4. When Codex needs access, approve it
5. Let it keep executing in the background

Typical Usage

@Computer Play some music to help me focus.
@Computer Please add my interview notes from Notes to my Feishu doc.
@Computer Please check my WeCom and set reminders for everything I need to finish before the end of today.

Caveats

• Make sure the Mac won't lock while it runs, or enable the "operate while locked" option in Codex
• Don't run two Computer Use threads controlling the same app
• After each thread ends, you can ask Codex to summarize and refine, turning the workflow into a repeatable pattern

/goal: Give Codex a Long-Running Objective

Normally, getting work out of an AI means standing over its shoulder. /goal solves exactly that: give Codex a long-term objective and let it keep working toward that direction on its own.

When to Use It

• Project migrations: moving a game to a new tech stack, a mobile app to a new platform, a codebase to a new framework
• Prototyping: building a new app or feature from scratch, with PLAN.md spelling the goal out clearly
• Prompt tuning: with a test set in hand, have Codex study failure cases, revise the prompt, rerun the eval, and iterate until the score climbs

How to Write a Good Goal

1. Give one clear objective and one done condition
2. Tell it which files, docs, or plans to read first
3. Fix which command will prove progress
4. Have it work in stages, keeping a brief progress log along the way
5. Check status anytime with /goal
6. Pause, resume, or clear it when the run finishes, gets stuck, or needs a new direction

$$slides + $$imagegen: Making Slide Decks with Codex

The most soul-grinding part of deck-making is layout. Codex ships with two skills: $$slides reads and writes .pptx files directly via PptxGenJS, and $$imagegen handles generating the artwork.

Reference Prompt

Use the $$slides and $$imagegen skills to edit this deck as follows:

- If present, add the logo.png file to the bottom-right corner of every slide
- On slides X, Y, and Z, shift the text to the left and use image generation to create an illustration on the right (style: abstract, digital art)
- Keep text as text wherever possible, and simple charts as native PowerPoint charts
- Render the updated presentation to slide images, inspect the output, and fix layout issues before delivering
- When creating a batch of related images, save the reusable prompts or generation instructions

Building Web Pages from Screenshots

Hand Codex a screenshot or design mock and it can build a responsive interface from it, while following the project's existing framework and conventions instead of starting its own. Paired with $playwright, it can compare the page against the screenshot across different screen sizes in a real browser and adjust repeatedly.

Reference Prompt

Implement this UI in the current project based on the screenshots and annotations I provide.

Requirements:
- Reuse the existing design system components and markup
- Map the screenshot onto this repository's utility and component patterns
- Match spacing, layout, hierarchy, and responsive behavior closely
- Respect the repo's routing, state, and data-fetching patterns
- Make the page responsive on both desktop and mobile

Verification:
- Use $playwright-interactive to check whether the UI matches the reference, and iterate as needed

Building Games with Codex

Games are the scenario that stress-tests everything Codex can do. Before writing any code, have it produce two documents:

PLAN.md -- break the game down into: player goals, core loop, controls, win/lose conditions, difficulty growth, visual direction, tech stack, and milestone order.

AGENTS.md -- spell out the tech stack, conventions, and how to verify.

Game name
<game genre>

Tech stack:
- Frontend NextJS (deployed on Vercel)
- Rendering with <fill in tech>
- Backend Fastify + WebSocket
- Database Postgres

Conventions:
- Verify every finished feature with the build / test commands
- Follow PLAN.md when building new features
- Use playwright to test the visuals
- Use imagegen for assets, storing prompts into .prompts

Install every skill AGENTS.md mentions: $$imagegen, $$playwright, $openai-docs, plus Context7 MCP to pull rendering-framework docs if needed. The more detailed the plan, the more presentable the first version comes out.

Running Research with Codex

Codex can serve as a long-term research assistant. You supply the direction and the judgment; it implements, gathers evidence, scores, and iterates.

Typical case: with a protein-folding hypothesis in hand, Goal Mode gives it three things — a scientific direction, a runnable baseline model, and an auto-scoring benchmark. In the official case, Codex ran for more than 150 hours straight and produced an experimental architecture called SimplexFold.

FAQ

• Where do I open Codex? In the ChatGPT sidebar, available to Pro/Team/Enterprise users
• Does it consume a lot of tokens? Far more than ordinary conversations; start with small tasks
• Which apps does Computer Use support? In theory all desktop apps, driven by UI element recognition
• Is Codex about to get cheaper? According to foreign media, OpenAI is considering steep fee cuts to compete for customers

More cases: developers.openai.com/codex/use-cases