Getting started with Codex

You just started using Codex. Maybe this is your first AI coding tool, or you swapped over from Claude Code, or you just want to give it a shot to compare AI tools.

Here's a guide to getting started with Codex so that you can start using it effectively ASAP. These are all tips from my own experience using Codex for 12+ hours everyday!

Codex CLI vs Codex App

There's 2 main ways to use Codex to code - via the CLI or via the Codex app. It's mostly up to personal preference so I recommend trying both, but here's how I use them:

• Codex app - I always have this open so I can start up chats in multiple projects at once (even just to get the thread/planning started). Also comes with automations and a nicer UI for plugins/skills. The Codex app also allows you to work with worktrees + cloud agents too, which is a pretty big bonus.

• Codex CLI - if you prefer CLIs then definitely use this. I like using this if I'm using a terminal (e.g. reviewing changes in Lazygit) and want to quickly spin up a chat to do something quickly.

An important note is that they're releasing a unified app soon that will have chat + coding + browser + more integrated.

What do I configure first?

I'll break this down into 2 sections because it's different for the CLI and app.

Codex app (CMD + ,)

• The most important setting is the theme - go to Appearance and set your theme up.

• Personalization - change Codex's personality and add custom instructions here. Note that "Custom instructions" is just your AGENTS md file from ~/.codex

• General - set "Speed" to "Fast" if you want 1.5x speed for 2x usage. I turn this on because I don't go crazy with subagents so I have spare usage. I also like "follow-up behavior" as "Queue" since I like queueing up commands, but you'll see which one you prefer as you use it. Adjust "Notifications" settings based on if you like being notified when the agent is finished or not.

Codex CLI (~/.codex)

• Create/update the AGENTS.md file. When I started using Codex I just looked at Peter's one here and adjusted as I used Codex more -

  https://github.com/steipete/agent-scripts/blob/main/AGENTS.MD

  . To edit personality, run /personality.

• When you're running Codex, run /experimental. This shows you beta features that you can enable. I usually turn them on whenever there's a new one because it's usually useful, but I have "Guardian Approvals" on because I run Codex in yolo mode.

• If you want fast mode (1.5x speed at 2x usage), run /fast.

Which model do I use?

Most people use GPT 5.4, but the main decision comes down to which reasoning level you prefer. I pretty much always use High, but I've seen

@steipete

recommend medium as well, so either one works. A lot of people have different recommendations here so try them all out. I personally never use "Extra high" because it takes too long and costs more for not much of an improvement over "High".

Which permission settings do I use?

I think most people run their Codex on "Full Access", probably because Codex used to ask for permission way too often and the permission settings weren't that configurable.

If you're not that comfortable with full access, you can use "Guardian Approvals" in the CLI which is a security reviewer subagent that reviews risky commands. In the app in Settings -> Configuration, you can also adjust permissions based on your preferences. Here's an article with a more detailed breakdown:

https://developers.openai.com/codex/agent-approvals-security

Which skills/MCPs/plugins/slash commands/automations do I configure?

Nothing. Don't use anyone else's prebuilt Codex setup. AI workflows and configs are HEAVILY dependent on the individual.

Just start using Codex in your day-to-day work and start adding things on slowly that way, so that your setup is optimized for your usage. E.g. if you repeat something everyday, add a slash command. If you do something every few hours, create an automation.

General tips

• Make sure the wording in your prompt isn't vague because Codex may interpret it incorrectly. It's a little bit worse than Claude in this regard.

• Codex excels extremely well at tasks with an acceptance criteria/verification method, even if it's something that takes hours.

• It's not the greatest at frontend at the moment so I'd just use a different model for now, but this is actively being worked on so expect improvements soon.

• Use phrasing like "don't stop until XYZ" to get Codex to run for a long time to finish something.

• Try to prevent over-bloating your AGENTS.md file. Codex is really good at following instructions, which can sometimes be a double edged sword if your AGENTS.md is bloated. Make sure it's clean so it has clear instructions and guidelines.

• Codex invokes skills very often which can be both a good and bad thing. It's good because you don't need manual invocation that much, but it's bad if you have way too many useless skills - it'll start invoking them out of nowhere. That's why I recommended setting nothing up at first.

In conclusion

You don't need to configure too much - just configure the basics like permissions and your AGENTS.md, and then go from there. If you start out and try to configure 50 subagents and 25 skills, you'll get overwhelmed and they won't be tailored to your day-to-day work.

If you have any questions feel free to reply or email me.

As the OpenAI team always says - you can just build things.