5.4 Your Codex Power-User Checklist

One condensed reference, worth pinning somewhere you'll actually see it again.

Every new project:

Generate an AGENTS.md, then hand-edit it — commands, architecture, conventions, guardrails, stated exactly

Set your default sandbox and approval policy deliberately, not by leaving whatever was default

Check which MCP servers are relevant here and connect them once, up front, instead of re-explaining context all project long

Every session:

Know your sandbox and approval settings before you start, not after something happens you didn't expect

Plan mode for anything unfamiliar or genuinely risky

Delegate research and exploration to subagents to keep your main context clean for the decisions that actually need you

/diff before you consider any set of changes "done"

/compact at a natural breakpoint, deliberately — not after you're already lagging

Every week:

Check which model your sessions are actually running

Skim OpenAI's Codex release notes for what's new

Notice any task you've described the same way twice — that's your next Skill or AGENTS.md addition

Before shipping anything real:

Run a Codex Security pass on anything genuinely sensitive

Confirm nothing leaked and no destructive command ran outside the scope you intended

Update AGENTS.md, or write a Skill, for anything you already know you'll repeat

That's the course. You now know more about what Codex actually does than most people who use it every single day without ever going past pasting snippets from a chat window. The only thing left is reps — pick something real, and go build it.