AI Tools · Oct 8, 2026
How to Use OpenAI Codex: A Beginner's Walkthrough From Prompt to Working Game
By ChannelGlance Editorial · Approx. 5 min read

OpenAI Codex can take a single plain-English prompt and return a working software project, and Kevin Stratvert’s step-by-step demonstration shows the full beginner arc on one example: from an empty folder, a request for a cookie-shop tycoon browser game produced a playable build in roughly three minutes and ten seconds, with no code written by hand. The same walkthrough then layers on the habits that make an AI coding agent practical — reviewing file diffs, saving Git checkpoints without commands, running three agents in parallel, packaging tests into a reusable skill, and delegating a feature to Codex Cloud through GitHub. This article explains each step and the reasoning behind it, so you can follow the workflow without watching anything.
Set Up the ChatGPT Desktop App and a Project Folder
The starting point is deliberately small. Install the ChatGPT desktop app on Windows or Mac, select Codex from the top-left of the interface, and create a project. A project is tied to a folder on your computer, and that folder is where Codex reads and writes files.
The desktop app is not the only door in. Stratvert also surveys the other access points:
- a web interface you can use from a browser
- a command-line tool for terminal-first users
- a VS Code extension for people who already live in that editor
The underlying workflow of describing tasks and reviewing changes stays the same across these surfaces. For beginners, the desktop app keeps everything in one window, with the game visibly running alongside the conversation.
Prompt a Working Browser Game in Plain English
With the project open, you describe what you want the way you would describe it to a person. In the demo, one sentence asking for a cookie-shop tycoon game was enough for Codex to create the files and produce something playable in about three minutes and ten seconds. Treat that number as an observation from a single demonstration, not a promise; your results depend on the project and the machine.
Iteration works the same way. After playing the game, Stratvert asked for animated customers, and the change arrived in just under two minutes. The pattern worth copying is simple: prompt, try the result yourself, then describe the next improvement in ordinary language rather than editing files by hand.
Review File Diffs Before You Keep a Change
Every change Codex makes can be inspected as a diff — a per-file view with additions and removals, shown in the demonstration as green and red lines. This matters for two reasons. First, it keeps you in control: you can see what the agent touched instead of accepting a black-box result. Second, it is a gentle way for non-coders to learn which files exist and how small edits connect to visible behavior.
One honest limit: the demonstration shows diffs working as intended but stops short of critiquing code quality, so it cannot tell you whether the generated code was clean, efficient, or maintainable. Reviewing diffs yourself is the workflow’s built-in answer to that gap.
Create Git Checkpoints Without the Command Line
Codex integrates Git so you never open a terminal. You can view the changes each turn produced, and when a version works, you simply ask Codex to commit it. Each commit acts as a restore point: if a later change breaks something, you have a known-good state to fall back to.
For a beginner, this reframes version control from a memorized command set into a conversational habit — ship an improvement, test it, then say ‘commit this.’ Stratvert demonstrates the loop on the game project, saving working states as checkpoints before attempting anything ambitious.
Run Parallel Agents in Isolated Worktrees
When several improvements are independent, Codex can pursue them at the same time. The demo delegates three separate features to three agents, each working in its own Git worktree — an isolated copy of the project where changes cannot collide. The batch finished in about five minutes.
The workflow then rewards hands-on judgment. You open each candidate version, play it, and decide which to keep. Stratvert merged one feature back into the main project and let the others go. Parallelism saves waiting, but selective merging remains a human decision based on which result actually works.
Package Testing Into a Reusable ‘Game Tester’ Skill
Asking an agent to re-check the same things every session gets repetitive, so Codex supports skills: reusable instruction sets you can rerun on demand. The demo builds a ‘Game Tester’ skill that verifies:
- the game launches
- the controls respond
- the calculations are right
- invalid spending is rejected
Once saved, the skill turns repeated testing into a single request. This is the most transferable idea in the walkthrough — anything you find yourself re-explaining, such as launch checks or validation rules, can become a named skill applied to future versions of the project.
Hand Tasks to Codex Cloud Through GitHub
The final step moves work off your machine. Stratvert creates a GitHub account, connects it to ChatGPT, pushes the project to a repository, and then assigns a customer-satisfaction feature to Codex Cloud, which builds and tests the change remotely. You hand off a task, step away, and later review the result as a diff and a tested build.
The GitHub connection also gives the project a remote backup and a standard place for cloud agents to work from.
What the Walkthrough Leaves Uncovered
The source is broad but deliberately shallow, and three gaps are worth naming. It does not discuss pricing or usage limits, so budget questions remain open. It does not unpack prompt craft in depth — the demo prompts work, but the reasoning behind them is not explored. And it displays output quality without critically evaluating the generated code, so maintainability goes untested.
None of this undermines the workflow; it defines the reader the workflow fits. Non-coders and beginners get a complete, repeatable path from idea to working game, with checkpoints, parallel attempts, automated testing, and cloud handoff along the way. Experienced developers who already use Git worktrees and GitHub daily will find little here beyond a surface tour. Stratvert’s original tutorial, ‘How to Use OpenAI Codex (Step-by-Step Tutorial),’ remains useful optional context if you want to watch each step performed live.
Frequently asked questions
Do I need to know how to code to use OpenAI Codex?
No. The demonstrated workflow relies on plain-English prompts, and Codex creates and modifies the project files itself. You still review diffs and decide which changes to keep, which is a reading task rather than a coding task.
How do Git checkpoints work in Codex without commands?
You view the changes each turn produced and ask Codex to commit work in conversation. Each commit becomes a restore point you can return to if a later change breaks the project, with no command-line knowledge required.
What are parallel agents and Git worktrees?
You can assign separate features to multiple Codex agents at once. Each agent works in an isolated Git worktree, a copy of the project where changes cannot interfere with each other. You then test each version and selectively merge the one you want.
What is a Codex skill like the Game Tester?
A skill is a reusable set of instructions Codex can rerun on demand. The demo's Game Tester checks that the game launches, controls respond, calculations are correct, and invalid spending is rejected, turning repeated testing into a single request.
What does the source tutorial not cover?
It omits pricing and usage limits, does not explore prompt craft in depth, and shows code output without critically evaluating its quality. It also only briefly surveys the web, CLI, and VS Code access points, so depth on those surfaces is limited.