Codex Tutorial For Beginners: The Complete Crash Course

Codex is not just ChatGPT for code

Codex is easy to underestimate because the interface looks familiar. You type a request into an input box, choose a model, and wait for a response. Underneath that simple interface, Codex is doing something different from a normal chatbot.

ChatGPT mostly answers. Codex can act. It can inspect files, create folders, edit documents, run local tools, open browser sessions, connect to apps, and verify the output of its own work. That makes it useful for far more than programming.

If you are a non-technical business owner, operator, analyst, or manager, the important point is this: Codex does not require you to write code. It gives an AI model a controlled workspace where it can complete multi-step computer tasks on your behalf.

This guide covers the beginner workflow shown in the video: installing Codex, setting up a project, organizing business files, generating reports, using memory, creating reusable skills, scheduling work, connecting plugins, and extending Codex with MCP servers.

The best beginner setup is the standalone app

There are three common ways to use Codex:

• the standalone Codex app;
• an editor integration;
• the terminal.

For beginners, the standalone app is the right place to start. It gives you the full Codex workflow without forcing you into developer tools. You can create projects, approve actions, inspect work, connect apps, and manage scheduled tasks from a more approachable interface.

Codex is separate from ChatGPT as an app, but usage is tied to your ChatGPT subscription. After downloading it for macOS or Windows, sign in with your account and start with the default model unless you have a specific reason to change it.

Tune reasoning and approvals for the task

Codex lets you control how much the model reasons before and during a task. More reasoning can help with complex work, but it usually makes responses slower. For ordinary business tasks, light reasoning is often enough. If the model struggles, increase the reasoning level and retry the task.

The second important setting is approvals. Codex can interact with your computer, so it needs a permission model. By default, it may ask before many actions. That is safe, but it can become tedious when a single task requires many harmless file reads or edits.

For normal everyday work, an approval mode that only interrupts for potentially unsafe actions is usually the practical choice. Full access should be reserved for trusted contexts where you understand the risk.

You can also switch the work mode from coding-focused responses to everyday work. That makes Codex less technical in its explanations, which is useful when you are asking it to handle documents, spreadsheets, reports, research, or operations tasks.

Put Codex inside a project folder

The most important beginner habit is to create a project folder before asking Codex to act.

A project gives Codex a bounded workspace. If you create a folder called Codex1 and select it as the project, Codex can work inside that folder without roaming across the rest of your computer. That containment is useful for safety and clarity.

For example, you can put a set of dummy business files inside the folder: accounting documents, delivery notes, client records, invoices, and reporting files. Then you can ask Codex:

Please organize my files into subfolders and give me a P&L segmented by client.

Codex can inspect the folder, infer categories, create subfolders, move files, and produce a spreadsheet with the resulting profit and loss analysis. It can then open or preview the spreadsheet so you can review the output.

That is the basic pattern: give Codex a contained workspace, describe the outcome in clear English, and let it use tools to complete the work.

The agent loop is what makes Codex different

The reason Codex can complete multi-step work is the agent loop.

In a normal chatbot flow, you send a prompt, the model generates a response, and the interaction ends. Codex can instead decide to use a tool. The result of that tool call goes back into the model, and the model decides what to do next. It may read more files, create a document, run a check, take a screenshot, or finish with a final response.

That loop matters because real business tasks are not one-shot text completions. They require inspection, action, feedback, and correction.

A useful Codex task often looks like this:

1. Understand the user’s request.
2. Inspect the relevant project files.
3. Decide what artifacts need to be created or changed.
4. Execute the file, browser, app, or terminal actions.
5. Verify the output.
6. Report what changed and where to find it.

That is why Codex is useful for work such as organizing files, generating reports, creating dashboards, preparing invoices, summarizing email, or turning source documents into structured outputs.

Reference files directly in your prompt

Codex supports file references inside prompts. In the app, you can use the @ symbol to reference a specific file in the project.

Instead of saying “use the spreadsheet in this folder,” you can point Codex to the exact file, such as a client P&L spreadsheet. Then you can ask:

Create an interactive dashboard for @Client P&L Q1 2025.

That reduces ambiguity. Codex knows which file to use, and the referenced file becomes part of the task context.

This is useful whenever the folder contains several similar files. It is also useful when you want Codex to avoid scanning unrelated documents. The more precisely you point to the inputs, the less time Codex wastes finding them.

Prompting Codex should be boring and specific

You do not need elaborate prompt formulas. In most cases, the best prompt is clear English that says what you want, what inputs to use, and what finished output should look like.

Avoid wasting tokens on generic roleplay such as “you are an expert data analyst.” Modern models already understand common professional tasks. Extra boilerplate can dilute the context without improving the result.

For complex tasks, add verification instructions. For example:

Create PDF invoices from this spreadsheet. After you are done, open a few generated PDFs at random and confirm that the client name, amount, and due date are present.

That is a useful instruction because it changes the workflow. You are not asking the model to sound more confident. You are asking it to inspect its output.

If your prompt is long, put the most important details near the beginning. Long context windows are useful, but models can still miss details buried in the middle of a large prompt. State the goal, constraints, and success criteria early.

Use memory intentionally

Codex has two kinds of memory:

• global memory;
• project-level memory.

Global memory lives in Codex settings. It can include personality preferences, custom instructions, and remembered context. Some users like it because it makes Codex adapt across projects. Others prefer to keep it off so every prompt stays explicit and controlled.

Project-level memory is more predictable. You create a file named AGENTS.md inside a project folder, and Codex reads it as instructions for that project. The filename matters. It needs to be exactly AGENTS.md.

An AGENTS.md file can include rules such as:

• how files should be named;
• what format reports should use;
• which folders are source material and which folders are outputs;
• what accounting categories should be used;
• what quality checks Codex should run before finishing.

This is a strong pattern for recurring business workflows. Instead of repeating your rules in every prompt, put stable instructions in the project.

Skills are reusable prompts

Skills are reusable instructions that Codex can invoke when you want it to perform a familiar workflow. Think of a skill as a saved operating procedure.

For example, you can create a skill that generates square social media images using your brand colors. Once the skill exists, you can invoke it later instead of pasting the same creative rules, dimensions, colors, and formatting requirements every time.

The practical beginner workflow is:

1. Leave the current project if you are creating a global skill.
2. Invoke the skill creator.
3. Describe the repeatable workflow.
4. Include stable rules such as brand colors, output size, tone, and file format.
5. Inspect the generated skill.
6. Test it inside a real project.

Skills are useful when the same instruction pattern keeps coming back. They are not necessary for one-off work. Use them when repetition is starting to cost time or create inconsistency.

Scheduled tasks turn prompts into recurring automation

Scheduled tasks let Codex run a prompt at a recurring interval. They used to be called automations, and they are one of the most practical ways to move from “AI assistant” to “AI operations worker.”

A scheduled task can run daily, weekly, on weekdays, or on a custom schedule. It can run globally or inside a specific project. For non-code work, running locally inside a project is often the right choice.

For example, you can create a daily task:

If today is a holiday, use the branded social media image skill to create an image for our social media.

If the day is a holiday, Codex creates the image. If not, it does nothing. The same pattern can apply to business workflows:

• summarize recent emails every morning;
• create weekly accounting reports;
• follow up with leads;
• collect competitor updates;
• prepare recurring social posts;
• check a folder for new files and process them.

The value is not that Codex can answer a question. The value is that it can repeatedly run a defined process.

Plugins connect Codex to apps

Plugins let Codex interact with other apps and websites. Gmail is a straightforward example. Once connected, Codex can search emails, draft messages, summarize threads, or combine email data with a scheduled task.

The important habit is to give Codex a clear goal and a clear boundary. For example:

Use Gmail to find new inbound leads from the last 24 hours and prepare a short summary. Do not send any replies.

That kind of prompt makes the expected behavior clear. It also prevents Codex from taking actions you did not intend.

Two of the most flexible plugins are Computer Use and Chrome. Dedicated plugins are cleaner when they exist, but many apps do not have a purpose-built integration. Computer Use can operate desktop apps, and Chrome can operate websites.

Use those carefully. They are powerful because they can interact with almost anything, but that also means you should keep prompts specific and inspect results before relying on them in production workflows.

MCP servers expand the tool ecosystem

MCP servers are similar to plugins, but they are not specific to Codex. MCP is a general interface that lets AI systems connect to external tools and services.

That matters because an MCP server can often work across several AI environments, including Codex and other coding or agent tools. It also gives access to third-party MCP marketplaces. Apify, for example, offers a large collection of tools that can be connected through MCP.

For a business user, MCP servers are useful when you need Codex to interact with a specialized service that is not available as a native plugin.

Typical use cases include:

• scraping or monitoring public websites;
• connecting to internal business systems;
• querying custom APIs;
• enriching leads;
• moving data between tools;
• running specialized data extraction workflows.

Start with built-in plugins when they cover the need. Reach for MCP when you need a broader or more specialized integration.

Codex can also work from your phone

If you have the ChatGPT mobile app, you can send prompts from your phone to your computer. The devices do not need to be on the same Wi-Fi network.

This is useful for lightweight task delegation. If you are waiting somewhere and remember a task, you can send it to Codex instead of waiting until you are back at your desk.

The best mobile prompts are bounded and low-risk:

• summarize a folder;
• prepare a draft;
• inspect a file;
• generate a report;
• collect information for review later.

Avoid initiating risky actions from mobile unless you are prepared to review the result carefully.

Advanced users can edit the Codex config

Codex stores configuration files locally. On macOS, you can open the config folder from the terminal, then inspect files such as the global AGENTS.md and config.toml.

The global AGENTS.md contains your global custom instructions. The config.toml file can include settings such as the selected model, personality, and reasoning level.

Advanced users can also configure Codex to use other model providers through services such as OpenRouter. That makes it possible to use Codex with models outside the default OpenAI setup, including Claude or open-source models exposed through a gateway.

This is optional. Beginners should not start here. The normal app settings are enough for most workflows. Config editing becomes useful when you want consistent global behavior across projects or want to experiment with custom model routing.

Where Codex fits in a business

Codex is best understood as an AI workbench for computer tasks. It is not limited to software development, and it is not magic. It works well when you give it:

• a bounded project folder;
• clear instructions;
• specific source files;
• stable project rules;
• permission to use the tools needed for the task;
• verification criteria.

It struggles when the task is vague, the files are messy, the desired output is undefined, or the workflow requires judgment that has not been explained.

For business teams, the opportunity is to identify repeatable work that already happens on a computer: reports, document generation, inbox summaries, spreadsheet cleanup, website checks, CRM updates, research briefs, and internal dashboards.

Start with one low-risk workflow. Put the files in a project folder. Add an AGENTS.md file with the rules. Ask Codex to perform the task once. Review the result. Then turn the stable parts into a skill or scheduled task.

That progression is how Codex becomes useful: not as a chatbot you occasionally ask questions, but as a controlled agent that handles repeatable computer work. If you want help identifying workflows, connecting tools, or building reliable AI automations for your business, you can book a call.