How to Use Codex (Even If You're Not a Developer)

Sources checked November 5, 2026: OpenAI's Codex quickstart, pricing, AGENTS.md guide, approvals and security, memories and skills docs. Plans, models and menus change often, so check those pages before you buy. For developer surfaces (CLI, VS Code, cloud tasks), see our full Codex setup guide.

What Codex is, in plain terms

Codex is OpenAI's coding agent. The word "coding" scares off people who would get the most from it. What makes Codex different from a chat window is that it works in a folder. You point it at a set of files, describe a job, and it opens what it needs, makes the changes, runs small scripts if the job calls for one, and reports back.

To Codex, a project is just a folder. That folder can hold code, but it can just as easily hold blog drafts, a CSV export from your CRM, campaign briefs or a pile of meeting notes. The agent doesn't care that the files aren't software.

Compare that with how most people use ChatGPT. You paste text in, get an answer, copy it back out to wherever it belongs, and repeat for the next file. Codex removes the shuttling. A task that touches twenty files, like rewriting twenty product descriptions or merging three spreadsheets, becomes one request.

Codex or ChatGPT Work: which one you want

This is the first thing to get right in late 2026, because OpenAI now offers two agent modes in the same app. According to its quickstart:

• ChatGPT Work is for "research, analysis, or deliverables such as documents, presentations, spreadsheets, and Sites."
• Codex is for "software development with codebase context and developer tools."

They share one pool of usage: OpenAI says Work inside ChatGPT "uses the same pricing, credits, and usage limits as Codex" (pricing). So the choice isn't about cost. It's about the shape of the job.

If you've been comparing Codex with Anthropic's agent instead, our Cowork vs Codex guide covers that choice.

Set up Codex without the terminal

You don't need the command line. Developers often use the Codex CLI or the IDE extension; you can skip both.

1. Install the ChatGPT desktop app and sign in

The desktop app runs on macOS, Windows and Linux. Sign in with your ChatGPT account (OpenAI). On a company Business or Enterprise workspace, your admin controls which features and plugins are switched on, so ask them if Codex doesn't appear.

2. Make one working folder

Create a folder just for this, something like marketing-work, with a few subfolders that match how you work: briefs, drafts, data, approved. Keeping Codex in one folder makes its changes easy to review and keeps it away from files it has no reason to touch.

3. Choose Codex and open the folder

In the app, select Codex from the ChatGPT dropdown, then start a chat or open a folder. OpenAI's quickstart is direct about what that means: it "can read and modify files in the folder you choose."

4. Check the permission level before the first real task

Locally, Codex runs inside a sandbox that usually limits it to your working folder, and network access is off by default. In the standard setup it can read, edit and run commands inside the folder, and it asks before going outside the folder or online (OpenAI). If you only want to plan or ask questions first, switch to read-only with the /permissions command.

5. Run a test you can't get wrong

Drop one document into the folder and ask: "Summarize notes.md in five bullets and save the summary as summary.md." If a new file appears with a sensible summary, your setup works.

Brief once: the AGENTS.md file

Most people use Codex the way they use a chatbot. They open a chat, explain the audience, explain the tone, list the words to avoid, then ask for the task. The next day they do it all again. That re-explaining is what we call Prompt Debt, and it compounds.

Codex has a built-in place for the standing rules: a plain text file called AGENTS.md. OpenAI's docs say Codex "reads AGENTS.md files before doing any work" (OpenAI). You don't mention it in your request. It's simply there in every session. This is the Brief-Once Workflow in its simplest form.

Step 1: Write down what you keep repeating

Open a blank file and write what you'd tell a new freelancer on day one: who reads your work, how you sound, words you never use, how headlines are formatted, what "finished" looks like. Be specific. "Friendly tone" tells Codex almost nothing. "Write like a senior marketer talking to a peer. No exclamation marks" gives it something to follow.

Step 2: Save it in the right place

Save it as AGENTS.md in the top level of your working folder. Codex also builds a chain of these files: a personal one in your Codex home folder applies to every project, and files in subfolders add narrower rules. Files closer to the folder you're working in come later in the combined instructions, so they win when rules conflict (OpenAI).

Step 3: Keep your requests short

With the rules handled, a request can be one line: "Draft a LinkedIn post from briefs/q4-launch.md." Codex already knows your voice, audience and formatting.

Step 4: Fix the file, not just the output

When Codex gets something wrong that you'd want avoided next time, add a rule. OpenAI recommends treating this as a feedback loop: correct the assumption in AGENTS.md and ask the agent to update the file so "future sessions inherit the fix" (OpenAI). After a few weeks, the corrections mostly stop.

Two limits worth knowing. Codex stops adding instruction files once they reach 32 KiB combined by default, and OpenAI's advice is to keep the file small (AGENTS.md guide). A short list of specific rules works better than a long essay.

A complete AGENTS.md you can copy

Here's a starting file for a small B2B marketing team. Replace the bracketed parts.

# AGENTS.md

## About us
We are the marketing team at [Company], which sells [product] to [audience].
Readers are busy professionals who know their own industry. Never explain basics.

## Voice
- Plain, direct English. Second person ("you") in posts and emails.
- Sound like an experienced colleague, not a brand. No hype.
- No exclamation marks. Never use "unlock", "leverage", "seamless" or "game-changer".
- Start with the point. No scene-setting openers.

## Formatting
- Headlines under 60 characters, sentence case.
- Every blog post ends with an FAQ. Never add a "Conclusion" heading.
- Emails: three short paragraphs at most, no bullet points.

## Facts
- Never invent statistics, prices, customer names or quotes.
- If a fact needs checking, write [VERIFY] next to it instead of guessing.
- Every statistic in a blog post needs a source link.

## Files
- Briefs live in /briefs. Read the matching brief before drafting.
- Save drafts in /drafts with the date first: 2026-11-05-launch-post.md.
- Never edit or delete anything in /approved.

## When you finish
- List the files you created or changed.
- Flag anything you were unsure about.

Some judgment doesn't fit in a rules file: how a seasoned SEO lead would structure a page, or how a sales director would tear apart a cold email. That's where a Callstand comes in. Codex can connect to tools over MCP (OpenAI), so it can call an expert's packaged judgment when a task needs it. Your AGENTS.md covers house rules; the expert covers the calls you'd otherwise hire a consultant for. For the difference between that kind of judgment and plain instructions, see The Judgment Layer.

Seven first tasks for non-developers

The best first tasks touch several files or would cost you an afternoon of copy and paste. Each request below works as written once your folder and AGENTS.md are in place.

1. Rewrite old content in your current voice. "Rewrite every file in /blog-archive to follow our voice rules. Save the new versions in /drafts and leave the originals untouched."
2. Clean a messy export. "Open data/leads-export.csv. Remove duplicate rows, standardize company names and save data/leads-clean.csv. Tell me how many rows you removed."
3. Turn meeting notes into actions. "Read every file in /notes from this week and create actions.md listing each action item with the owner and due date mentioned."
4. Build a comparison from saved research. "Using the competitor pages saved in /research, create a table of pricing models and main features. Mark anything you can't confirm from the files with [VERIFY]."
5. Draft from a brief. "Draft a blog post from briefs/inventory-guide.md. Follow the outline exactly and save it in /drafts."
6. Check a batch against your rules. "Review every file in /drafts against AGENTS.md. Don't edit anything. Write review.md listing each rule broken, by file."
7. Make a small internal tool. "Create one HTML file called roi-calculator.html where someone enters monthly ad spend and conversion rate and sees cost per lead. No external libraries."

The last one surprises most non-developers. Codex writes the code; you open the file in your browser and use it.

How to review what Codex did

Reading the changes is the most important habit you can build. Codex will occasionally touch something you didn't ask it to.

• Ask why. "Why did you change the second paragraph of drafts/launch-post.md?" The answer often points to a vague rule you can tighten.
• Check facts yourself. Codex is only as accurate as the files you give it, and it can still fill gaps with confident guesses. The [VERIFY] rule helps it flag uncertainty instead of hiding it.
• Correct precisely. "Cut the intro to two sentences and move the pricing mention to the end" gets you further than "make it better."
• Promote repeat corrections. If you give the same correction twice, it belongs in AGENTS.md.

Skills, memories and connectors

Once the basics work, three features extend what Codex does without making you write code. OpenAI describes them as complementary layers alongside AGENTS.md (customization overview):

• Memories carry useful context from earlier work into later work. OpenAI is explicit that required rules still belong in AGENTS.md, with memories as "a helpful recall layer" (OpenAI).
• Skills package instructions, resources and optional scripts so Codex can follow a repeatable workflow, built on the open agent skills standard (OpenAI). A monthly report process is a good candidate. Our explainer on agent skills covers the format.
• MCP connections and plugins let Codex use tools and data from other systems, such as a ticketing tool or your docs (OpenAI). If you're unsure how that differs from an API, read MCP vs API.

The order matters. Start with AGENTS.md, add a skill when a multi-step process repeats, and add connectors when the files you need live somewhere else.

What it costs

Codex comes with ChatGPT plans rather than as a separate product. OpenAI's pricing page lists, as of November 5, 2026:

Usage isn't a fixed number of messages. OpenAI says model choice, context, reasoning and tool use all affect it, and that tasks that look similar can use different amounts. Its published estimate for Plus is 15 to 160 local messages per five hours on GPT-6.1 Sol. Bigger folders and longer sessions use more. Start on the plan you have, watch the usage dashboard for a week, then decide.

Frequently asked questions