Official References: Quickstart · CLI Slash Commands · Best Practices · Review
Curriculum path
- Codex Getting Started — install, sign in, first task, Git checkpoints ← You are here
- Codex Instructions —
AGENTS.md,/init, and config basics - Codex Sandboxing — permissions, approvals, and cloud environments
- Codex Task Design — issue-shaped prompts and task sizing
- Codex Skills — turn repeated prompts into reusable workflows
- Codex MCP — connect external context instead of pasting it
- Codex Reviews and Automations —
/review, worktrees, and repeatable engineering loops
Official docs used in this guide
- App / CLI / IDE starting flow — Quickstart
- Core commands like
/init,/review,/mcp, and/status— CLI Slash Commands - Why Git checkpoints matter early — Quickstart
Start with a success loop, not with every feature
The fastest way to get comfortable with Codex is not to memorize every setting. It is to complete one small, safe task end to end.
Early on, you only need to answer four questions:
- Which client will you use?
- Which repository will you open?
- How small is your first task?
- How will you roll back if you dislike the result?
Where should you start?
Codex can be used through the app, CLI, IDE extension, and cloud flows.
Codex app
Best for most people at the start. Review, automations, and worktrees are visible in one place.
CLI
Great if you live in the terminal. It is the clearest way to see prompts, commands, diff generation, and verification loops.
IDE extension
Useful if you prefer to stay inside your editor. Still, for your first few sessions, the app or CLI usually makes the behavior easier to understand.
Cloud
Do not start here unless you already know what your local workflow should look like. Local success first, cloud later, is much less confusing.
Your first 10 minutes
The Quickstart flow is straightforward:
- Install the Codex app or CLI
- Sign in with your ChatGPT account or an OpenAI API key
- Select the project folder
- Send your first task
The docs note that signing in with an API key can limit some features, such as certain cloud capabilities. For most new users, local project + normal account sign-in is the simplest path.
Start local before you start ambitious
If you try to learn cloud environments, automations, wide permissions, and prompt design at the same time, it becomes hard to tell what actually failed.
Better first tasks:
- “Explain this repo in five bullets.”
- “Find the files most relevant to this error message.”
- “Identify dead-code candidates in this folder.”
- “Make a tiny refactor without changing behavior.”
That means: choose work that is easy to inspect and easy to revert.
Use Git checkpoints early
Quickstart explicitly recommends Git checkpoints. That is a good habit because Codex can change a lot quickly.
A beginner-friendly loop looks like this:
git status
git add -A && git commit -m "checkpoint: before codex task"If the result looks good, make another checkpoint after the task. This removes a lot of “I think this is better, but I can’t see what happened” anxiety.
Four commands to learn first
The slash command docs expose many commands, but early on you mostly need these:
/init
Generate an AGENTS.md scaffold in the current directory.
Learn this as soon as you get tired of re-explaining repo rules.
/review
Ask Codex to review your working tree. It is the easiest second pass after implementation.
/mcp
See the MCP tools currently available in the session. Useful when you need external docs, trackers, or internal systems.
/status
Check your current model, approval policy, writable roots, and session state. When Codex behaves differently than you expected, start here.
Your first prompt should feel like work, not like a wish
Avoid “make the whole app better.” Try something shaped like this instead:
Goal:
Find unused utility candidates under src/lib/.
Constraints:
- Investigation only; do not modify files
- Show import or usage evidence for each candidate
Done when:
- You list the top three candidates with reasonsThis gives Codex a target and a finish line.
A good first-week checklist
In week one, this progression works well:
- Small read-only local tasks
- Small code changes plus lint/test runs
- Use
/reviewto inspect the diff again - Use
/initto draft anAGENTS.mdfor the repo - Promote repeated work into skills or automations only after the workflow feels stable
The healthy order is task → review → rule-making, not “turn on everything.”
Fast summary
- Start with the app or CLI on a local repo.
- Pick a first task that is small and reversible.
- Use Git checkpoints from day one.
- Learn
/init,/review,/mcp, and/statusearly. - Go deeper with instructions, sandboxing, and task design only after your first loop feels solid.