learn-codex-with-phoebe / Session 1 of 6
Learn Codex with Phoebe · Session 1 of 6

Meet Codex: your AI engineering teammate

What Codex actually is, how to set it up on a regular ChatGPT Plus subscription, and the one prompting structure that makes it work. By Session 6, you'll have built and shipped a real personal website with it - starting today.

🟢 Easiest Builders Non-engineers welcome 45 min live + self-study
0-3 · Welcome 3-18 · Concepts 18-40 · Setup & first run 40-45 · Q&A
Part 0

Why this course exists

Everyone has a half-finished personal website in their head: an about page, the things you've built, the ideas you keep meaning to write down. This course gets it out of your head and onto the internet - built from scratch WITH Codex, OpenAI's coding agent, on nothing more than a personal ChatGPT Plus subscription. Six sessions, one real shipped website, and a working command of the tool by the end.

Live - presented in session Self-study - read after class ★ Build-along demo The project: YOUR website
★ What you walk out with today Codex installed and signed in on your machine, your project folder created, a site plan written by you and Codex together - and the GCOB prompting structure you'll use in every session after this one.
Part 1 · covers "intro to codex" + "codex 101"

What Codex actually is 6 min live

Codex is an agent: you give it a goal, it gathers context, takes action in your files, and produces something real - code written, tests run, changes reviewed. One agent, five doors to reach it.

The Codex agent goal → context → action → something real, verified 🖥️ Desktop app Codex mode on a folder ⌨️ Codex CLI the terminal - our home base 🧩 IDE extension inside VS Code ☁️ Codex cloud delegate tasks, work in parallel 🔍 Code review /review + GitHub PRs Same agent, same account, same AGENTS.md and skills everywhere - pick the door that fits the moment.
🔍 Click to zoom - one agent, five doors (this course lives mostly in the CLI)
LiveAn agent, not an autocomplete3 min

Codex is not a chat window that emits code snippets. It works like a junior engineer with a terminal: it reads your files, plans, writes code, RUNS it, reads the errors, fixes them, and shows you the result - looping until the goal is met or it needs your call. You review and steer; it does the labor.

  • It sees your project: point it at a folder and it explores the real files - no copy-pasting code into a chat box.
  • It acts under your permissions: what it may read, write, and execute is a setting you control (Part 2), not a hope.
  • It verifies: the habit that separates agents from autocomplete - it can run the thing it just built and check.
Real world

The moment it clicks for most people: you type "the contact link is broken on mobile" - and instead of an explanation, Codex opens the file, finds the overlapping element, fixes the CSS, and shows you a diff. Ten seconds. That gap between "telling you about code" and "doing the work" is the entire product.

LiveWhich door when - the five surfaces2 min
SurfaceReach for it whenIn this course
Codex CLI (terminal)Building in a local folder - full control, fastest loopSessions 1-5, home base
Desktop app (Codex mode)Same power, friendlier face - great if terminals feel hostileValid alternative throughout
IDE extensionYou live in VS Code and want Codex beside the editorOptional, Session 2 note
Codex cloudDelegate a task and close the laptop; run tasks in parallelSession 6
Code reviewCheck changes before they ship - /review locally, @codex on GitHub PRsSession 5
Non-engineer in the room? The desktop app's Codex mode does everything this course needs. Wherever a demo says "in the terminal", the app's chat box accepts the same words.
Self-studyModels, credits, and what Plus actually gets you3 min read

Codex is included in every ChatGPT plan - Free through Enterprise. On Plus ($20/month), usage is metered in 5-hour windows, and cloud tasks and code reviews share the same pool:

ModelCharacterPlus: messages per 5-hr window
GPT-5.6 SolDeepest reasoning - hard problems, gnarly bugs~15-90
GPT-5.6 TerraThe balanced default - most work happens here~20-110
GPT-5.6 LunaFast and light - quick edits, questions~50-280
GPT-5.4 miniBulk simple tasks~60-350
  • Switch anytime with /model - model plus reasoning effort per task.
  • The budget habit: default to Terra; reach for Sol when genuinely stuck; drop to Luna for small edits. Ranges are wide because message cost scales with how much thinking a task needs.
  • Hit the ceiling? Plus users can buy extra credits, or simply wait for the window to reset. For this course's website project, Plus limits are comfortable - a session's build fits well inside one window.
Part 2 · covers the quickstart + permissions docs

Setup on your Plus plan 5 min live

Two minutes of installing, one sign-in, one decision about permissions. That's the whole barrier to entry.

LiveInstall and sign in2 min · then Demo 1
★ Install the CLI (pick one)# macOS / Linux - official installer curl -fsSL https://chatgpt.com/codex/install.sh | sh # or via npm npm install -g @openai/codex # or Homebrew (macOS) brew install --cask codex # then, from any project folder: codex # → sign in with your ChatGPT account when prompted. Done.

Signing in with ChatGPT links Codex to your Plus plan - no API key, no billing setup, nothing else to configure. The desktop app route is even shorter: install the ChatGPT app, open a folder, pick Codex mode.

LivePermissions - the safety dial2 min

Codex acts under a permission gate you control with /permissions: from read-only (it can look, not touch) through auto-edit in the project folder, up to running commands with approval. Two rules carry you through this whole course:

  • Start where you're comfortable. Read-only on day one is a fine choice - promote it as trust builds. Approval prompts are a feature: every "may I run this?" is a chance to learn what it's doing.
  • The sandbox is real: Codex works inside your project folder's boundary; it asks before going beyond. Your website project is a perfect training ground - nothing in it can hurt anything.
The habit that matters When Codex asks to run a command you don't recognize, ask it "what does that command do?" before approving. It explains happily - and three of those explanations teach more terminal than most tutorials.
Self-studyThe other doors: desktop app modes and the IDE extension2 min read
  • Desktop app has three modes: Quick Chat (casual questions), Work (research, documents, analysis), and Codex (software work on a folder). Same account, same limits - mode just sets the working style.
  • IDE extension: Codex inside VS Code, seeing your open editor context. Same agent; useful once you're editing the site by hand alongside it (Session 4 territory).
  • Session continuity: codex resume picks up a previous conversation in the CLI - your project's context survives lunch.
  • Everything shares your setup: the AGENTS.md and skills you'll build in Session 3 apply across CLI, app, and IDE automatically.
Part 3 · covers the official prompting guide

Talking to Codex: the GCOB structure 4 min live

OpenAI's own guidance boils down to four elements. No syntax, no magic words - just four questions answered in plain language, then iteration.

Goal - what should exist when you're done? Context - what does it need to know? (files, examples, taste) Output - what form? (a page, a plan, a diff, a list of options) Boundaries - what must NOT change or happen? = one strong ask then iterate in follow-ups - results over steps, always. Start with the result you want, not the steps to get there. Codex plans the steps - that's its job.
🔍 Click to zoom - GCOB: the anatomy of a strong Codex ask
LiveGCOB in action, plus the Codex-specific habits4 min
Weak - steps, no goalMake an index.html, add a header tag, then some CSS, use flexbox for...
Strong - GCOBGoal: a clean one-page personal site skeleton I can grow into about/showcase/ideas sections. Context: I'm a data & AI leader building my personal site. I like editorial, minimal design - generous whitespace, one accent color, no template feel. Output: working HTML/CSS in this folder, plus a short note on how you structured it and why. Boundaries: no frameworks, no build step - plain HTML/CSS/JS I can read and learn from. Nothing goes online yet.

Three Codex-specific habits from the official guide, worth tattooing somewhere:

  • Bugs get reproduction steps: "the menu breaks when I resize below 400px" beats "menu is broken".
  • Name your constraints: "minimal changes", "don't touch the design", "keep it readable for a beginner" - boundaries prevent enthusiastic rewrites.
  • Ask for verification: end big asks with "then check it works and tell me how you verified". Agents that verify are agents you can trust.
Real world

A first-timer's actual session: her first prompt was "make my website" - she got a generic template and hated it. Rewritten with GCOB (real goal, her taste in context, plain-code boundary), the second attempt was recognizably HERS. Same tool, same evening - the four questions were the difference.

Part 4 · the running project

One website, six sessions 3 min live

Every session advances the same build. No toy exercises - by Session 5 this is live on the internet with your name on it.

1 · Setup 2 · Skeleton 3 · Standards 4 · Pages 5 · Ship 6 · Autopilot install, sign in, plan the site working homepage, the build loop AGENTS.md + skills: your taste, encoded about · showcase · ideas, polished GitHub Pages, live URL, /review habit cloud tasks, exec, MCP, maintenance Miss a session? Each page stands alone, and the repo state you need is one Codex prompt away: "catch this project up to session N".
🔍 Click to zoom - the roadmap: one website, six sessions
Demo 1 of 3

Install, sign in, first conversation ★ 8 min · everyone installs

Run the installer (Part 2 box) - or install the ChatGPT desktop app if you're going the no-terminal route.

Make the project folder: mkdir my-site && cd my-site - then run codex and sign in with your ChatGPT account.

First conversation, no stakes: "Introduce yourself - what can you do in this empty folder, and what do you need from me to be useful?" Read the answer; it's a better feature tour than any slide.

Run /status - see your model, permissions, and folder. Then /model - look at the options, leave it on the default (Terra).

Stuck on install? Pair with a neighbor and keep going on their machine - fix yours after the session. The point today is the conversation, not the plumbing.
Demo 2 of 3

Plan YOUR site with GCOB ★ 10 min · everyone plans

Two quiet minutes with pen and paper: what are your three sections? (Default: about / things I've built / ideas I'm chewing on. Rename them to sound like you.)

Write your first real GCOB prompt - use the Part 3 example as a skeleton, but the Goal and Context are yours. Ask for a PLAN, not code yet.

Run it: "...Output: write SITE_PLAN.md in this folder - sections, pages, what each contains, and the build order you'd suggest. Boundaries: plan only, no code yet, keep it under a page."

Discernment lap: read SITE_PLAN.md, disagree with one thing out loud, and tell Codex to revise it. The plan is yours, not its.

Approve the plan. That file is Session 2's starting gun.

Real world

The revision moment is where ownership happens: "swap the blog for a shorter 'ideas' section - I'll write 200-word notes, not essays" turned a generic plan into one a real person would actually maintain. Codex is a great planner and a better listener.

Demo 3 of 3

Feel the permission gate ★ 4 min · follow along

Ask Codex: "Create a README.md that describes this project in two sentences." Watch the approval prompt appear (unless you've already granted write access).

Before approving, ask: "What exactly will you write and where?" - see it explain itself. THEN approve.

Open /permissions and set the level you're comfortable keeping for this project. There's no wrong answer - only your answer.

After the session

This week ◐ 30 min total

Check yourself

Three questions before you go 🎯 ◐ 90 seconds

1 · What makes Codex different from a chat window that writes code snippets?

Agent, not autocomplete: the gap between "telling you about code" and "doing the work" is the whole product.

2 · The GCOB elements are...

Four plain-language questions from OpenAI's own guide - then iterate in follow-ups. Results over steps, always.

3 · Codex asks to run a command you don't recognize. Best move?

The approval prompt is a feature and a free terminal lesson. Understand, then decide - trust is built, not granted.

Source material

Official sources covered

This session teaches the working content of OpenAI's official onboarding materials - docs at learn.chatgpt.com and free courses at OpenAI Academy.

Codex quickstart (official docs)install, sign-in, first run - Parts 1-2 + Demo 1
Prompting guide (official docs)the GCOB structure + Codex-specific habits - Part 3
Models & pricing (official docs)Sol/Terra/Luna, Plus limits, credits - Part 1
Intro to Codex + Codex for Beginners (OpenAI Academy)agent concept, surfaces, onboarding - Parts 1-2
Permissions (official docs)working level here; full depth with sandbox details in Session 2

Session 1 cheat sheet · pin this

What Codex isAn agent: goal → context → action → verified result. Five doors, one agent - CLI is home base.
Installcurl -fsSL https://chatgpt.com/codex/install.sh | sh, then `codex` + sign in with ChatGPT. Done.
GCOBGoal · Context · Output · Boundaries. Results over steps. Iterate in follow-ups.
Key commands/status · /model · /permissions · /init (Session 2) · /review (Session 5) · codex resume.
Model habitTerra by default · Sol when stuck · Luna for quick edits. Plus meters per 5-hr window.
Trust habitUnknown command in approval prompt? Ask "what does that do?" first. Every prompt is a lesson.