Codex in one evening. A guide.

This guide moves you, in one evening, from "chat in a browser" mode to "AI as a working tool" mode. No code. No jargon.

1. Why leave the chat window at all

Chat is great for a quick question or a one-off draft. But you carry context into the window by hand, and then sort the result into files by hand too.

In Codex, the model can read a project folder, edit files, and run commands. You set the boundaries, review the changes, and decide what to keep.

One idea matters most: whoever is best at assembling context wins. Codex is a tool for assembling context.

2. What this is, in your terms

Simpler version: picture an assistant working inside one folder, who asks permission before anything risky and shows you every change. You describe the task in words and check the result.

3. Three mental switches

This is the hardest part. Not the install, but the habit change. Until you make these three switches, the tool will just feel like an awkward chat window.

1. From copy-paste to files. Stop copying answers out. Say instead: "rewrite this document," "create a file with this content," "fix this part." The model works with disk, not with your clipboard.

2. Context hygiene. The more you cram into one conversation, the worse the model holds on to the task. Rule of thumb: a new session per task. Keep anything permanent in files, not in chat history.

3. Check every single word. The model is mostly right, but wrong in places. A domain expert with AI replaces someone without domain expertise. The expert themselves is not replaced. You stay the editor, not the audience.

4. Setup in 15 minutes

1. Check the current access options in the official OpenAI docs. Codex is included in paid ChatGPT plans, but terms and limits change, so don't rely on old screenshots and prices.
2. Pick where to work. For a beginner the Codex desktop app is easiest: folders, files, and changes are right in front of you. There's also a CLI for the terminal, an IDE extension (VS Code, Cursor, and others), and a cloud version in the browser.
3. Sign in with your ChatGPT account. You don't need a separate API key to get started.
4. Create a project folder on your desktop. One folder, one project.
5. Open that folder in Codex and first ask it to list the files it can see.

A note on economics: Russian, Polish, and Ukrainian text uses noticeably more tokens than English. Limits run out faster, keep that in mind.

5. Your first session

Don't overthink it. Four steps:

1. Put all the context into the folder: a brief, notes, data, whatever you have on the project.
2. Open Codex in that folder.
3. Say, in plain words, what you need. For example: "read these files and put together a one-page brief from them" or "build a quick and dirty prototype from this spec."
4. Look at the changes before you accept them. Codex shows which files and lines it changed. That's your checkpoint.

If it's easier to explain the task out loud, dictate. Then reread it and check the result.

6. Permissions: how much freedom to give the agent

The key difference from chat: Codex doesn't just write text, it takes actions. So it has access modes, which you switch with the /permissions command.

• Read only. Codex looks at files and makes suggestions, but changes nothing without asking. Good for a first look at an unfamiliar folder.
• Work inside the project folder. The default: it reads, edits files, and runs commands inside the folder, and only goes outside it or to the internet after you say yes.
• Full access. No questions asked. Not for beginners, and not for a folder with anything valuable in it.

Beginner rule: start with the default mode and actually read what the agent asks you to approve. If you don't understand a request, ask the agent itself: "explain what this command will do."

One more safety net: at the very start, ask Codex to "turn this folder into a git repository and save a version after every step." Then any bad edit can be rolled back with one sentence.

7. Five moves that are enough for a beginner

You don't need to know a hundred commands. You need five.

1. "Read the folder and tell me what you found." This is how you check that the model actually sees your context.
2. "Create a file / fix this file." Basic work with disk.
3. "Search the web and put together a table." Research, right inside your workflow.
4. /review or "explain what you did and why." Have the model check its own changes. It's easier to catch mistakes that way.
5. /new. Finished a task - open a fresh session for the next one. Don't drag the clutter along.

A couple more for later: /model picks the model and how hard it thinks, /status shows the current session settings.

8. How to write better prompts

A prompt is what you type in. The quality of the answer equals the quality of the prompt. Garbage in, garbage out. This is the skill that improves fastest and pays off the most.

1. Role + task + constraints + format. Not "sort out these expenses," but "you're an experienced financial analyst, sort these expenses into categories, don't invent new ones, use mine, give me a table."
2. Give an example. Show 1-2 samples of what you want. The model copies a sample better than it interprets a description.
3. Say what NOT to do. "Don't use em dashes," "don't pad it with filler," "don't make up facts."
4. Plan first, then act. "Show me a plan before you start." You catch a mistake before the model spends time going the wrong direction.
5. Iterate. The first answer is a draft. "Too long, cut it in half," "add numbers," "rewrite it tighter."

The main trick: let Codex write prompts for you. "Write a prompt that would make you do this task well" or "improve my prompt: [your text]." Save a prompt that worked into a file in the project folder and reuse it. That's how a one-off success turns into a repeatable process.

9. AGENTS.md: rules the model reads every time

Think of it as a folder of documents. You build it by hand: project description, your role, the subject matter, market data, your target audience, the spec, marketing materials. The better the folder, the smarter the model.

One file is worth setting up right away: AGENTS.md at the root of the folder. Codex reads it at the start of every session. Who you are, what you're building, what the rules are (for example: "don't use em dashes, write in English, check facts, show a plan before editing"). Those are your guardrails.

Don't write it from scratch: the /init command creates a draft, and you add your own rules. A nice bonus: other AI tools understand AGENTS.md too, so your rules come with you if you switch tools.

Beyond this there's more advanced stuff: skills (ready-made capabilities for recurring tasks), MCP connections (email, calendar, databases, your own services), and cloud tasks Codex runs in the background while you do something else. That's what turns it from an assistant into a working system. We go deep on this in the course and in the practical group.

10. Codex or Claude Code?

A common question. Short answer: the approach is the same. A folder of context, a rules file, a session per task, a check of the result. Learn one, and you'll be at home in the other within an evening.

• Already paying for ChatGPT - start with Codex, no extra subscription needed.
• Already paying for Claude - start with Claude Code. There's a separate guide for it.
• Don't spend a week choosing. Pick either one and ship a first project. The skill of assembling context carries over; getting used to specific buttons takes a couple of days.

11. Beginner pitfalls

• Limits. Sooner or later you'll hit your plan's limit. That's normal, don't panic. The result is never perfect on the first try.
• Clicking "yes" without reading. Codex asks permission for a reason. Read what it's about to do.
• Context clutter. If the model starts getting dumb and forgetful - don't fight it, just open a new session.
• Blind trust. The most expensive mistake. Generate, then check. Especially numbers, facts, links.
• Vibe coding has a ceiling. A quick prototype is weak on quality, security, and scale. The moment it starts making money - rebuild it properly, with people who know what they're doing.
• Ten ideas at once. Pick one. Not three.

12. Where to start: simple projects

Don't invent a startup. Take a boring task from your own life and automate it. Pick one.

• Career agent. Resume, job listings, and notes in one folder. Ask it to: tailor your resume to a listing, find gaps in your experience, draft a cover letter.
• Finance agent. Export a bank statement (CSV or Excel): sort expenses into categories, find subscriptions, put together a monthly report.
• Telegram bot. A bot that answers common questions, collects requests, or sends reminders. Codex builds the skeleton, you connect it following instructions.
• Personal knowledge base. A folder of notes and write-ups. Ask it: "what did I write down about X," "pull everything about Y into one summary."
• Research assistant. Give it a topic: search the web, compare options, put together a table with pros, cons, and links.

Rule for picking: take whatever is bugging you personally right now. You're your own first user, and feedback is instant.

13. A quick win: from idea to prototype in one evening

1. Discovery. Segment your audience, sketch 3-5 personas, run a synthetic interview, research competitors. The outcome might be: nothing needs building, a solution already exists. That's a result too.
2. Solutioning. If the hypothesis holds up - write a light spec. One page: what problem you're solving, for whom, what makes you different.
3. Development. Put the spec in the folder and tell Codex: build it quick and dirty. Show it to 5-10 real people. Improve it or shelve it.

Think about go-to-market, pricing, and support only after value is confirmed. Not before.

This week's checklist

• ☐ Checked Codex access on your ChatGPT plan, installed the app.
• ☐ Picked one project. One.
• ☐ Created a folder, put all the context into it.
• ☐ Ran /init and added your own rules to AGENTS.md.
• ☐ Gave a first task in plain words and read the changes before accepting.
• ☐ Ran Discovery: personas, synthetic interview, competitor research.
• ☐ If the hypothesis held up - built a prototype and showed it to 5-10 people.

Whoever has more context wins. Start building yours.

← All AI Sandbox materials