Codex Technical Setup

Start with the boring tools. Install VS Code or another proper editor so the website is visible as files, not as scattered snippets in chat. Then install Git so changes can be reviewed and reversed with a history, not remembered from a conversation.

If the site uses a JavaScript stack, install Node.js and npm. For most non-technical users, the practical check is simple: open a terminal in the project folder and run node -v and npm -v. If both commands return versions, the basic runtime exists.

Do not install random packages because an AI answer sounds confident. Use the project's existing instructions first. If there are no instructions yet, create them before asking Codex to build anything substantial.

Codex can fit different comfort levels. The CLI is strong when you are comfortable in a terminal. The IDE extension is useful when you want Codex close to the editor. The Codex app is useful when you want a more guided place to manage agent work.

For a first website workflow, choose the surface that makes review easiest. The best setup is the one where you can see the changed files, read what Codex did, and run the project's checks before accepting the work.

If you use the CLI, the current OpenAI setup path is to install Codex from the official Codex CLI documentation, then run codex inside the project folder. The first run signs in with a ChatGPT account or API key. On Windows, use the native Windows setup, the IDE extension, or WSL2 depending on your project and company restrictions.

Create an AGENTS.md file in the project root. This is the file that tells Codex how this project works: what commands to run, what style to follow, what not to touch, where assets live, and what counts as done.

A useful first version can be short: project goal, install command, development command, build command, test or lint command, design rules, content rules, image rules, and any files that should not be edited without approval.

Keep this file practical. Codex does not need a brand manifesto when it is fixing a footer. It needs concrete instructions such as run npm run build before finalizing, use existing colors from the design file, and do not change tracking scripts unless the task asks for it.

Do not start with build the whole website. Start with a contained task that proves the workflow: create a reusable footer, adjust a header, add one content section, fix one mobile layout issue, or wire one tracking check.

A good first prompt has four parts: the goal, the relevant files, the constraints, and the done condition. For example: Update the footer to match AGENTS.md. Use existing components and colors. Do not change routing. Done when the page builds and the footer fits on mobile.

This is how you keep the AI useful. Codex can inspect files, edit code, and run commands, but you still define the box it should work inside. Small jobs make mistakes easier to see and corrections cheaper.

Every Codex task should end with evidence. At minimum, read the changed files, read the summary, and run the project's normal check command. For a website, that usually means a build command, a lint command if available, and a visual check in the browser.

If Codex says a command failed, treat that as useful information, not as a failure of the whole workflow. Ask it to explain the failure, fix the smallest cause, and rerun the same check. The goal is a repeatable loop: change, check, inspect, accept.

Never publish directly from an AI-generated change because it looks right in the conversation. Publish after the project checks pass and the visible page matches the expected outcome.

Once the basic setup works, turn repeated instructions into skills. A skill is useful when you find yourself giving Codex the same narrow workflow again and again: build a footer, create a thumbnail, prepare metadata, review mobile layout, or check tracking tags.

This is the bridge to the next guide. Codex is the workbench. Skills are the reusable tools on that workbench. Without Codex setup, skills are just instructions. With Codex setup, they become repeatable production workflows.