The Codex CLI course

You point OpenAI’s Codex CLI at a repo you did not write, and it is immediately more capable than you are at reading it and less trustworthy than the newest person on your team. Both halves of that have the same cause. Codex starts every turn knowing nothing about your codebase that it has not just read inside that same turn. It has no memory of yesterday’s session, no idea which conventions the code follows and which it violates, and no way to tell a rule from an accident.

That is the enemy this course attacks, and it does not go away. Every chapter here picks off one part of it, and every fix charges you something: tokens in the window, turns in the loop, blast radius when it is wrong, or maintenance when it rots. There is no free move. Learning which price to pay, for which job, is the whole skill.

The project is budgetcli, a small self-hosted personal-finance API you have just inherited, with your own money in it. You will spend a week turning it into something you would actually trust.

What the week is worth, in numbers

Onboarding a fourth bank’s CSV export takes about 11 turns on Monday, and somewhere in those turns the agent reintroduces the same money bug 3 separate times: it parses an amount as a float, in a codebase where every amount is an integer number of cents.

By Friday the same job takes 2 turns and the bug does not come back once.

The whole difference is 32 lines in one file, one saved profile, and one skill. Nothing exotic, and nothing you cannot build yourself in a week. The last chapter adds it up in front of you and shows which step took off what.

The four questions this course keeps asking

Every move you learn gets scored the same four ways, in a running ledger that grows a row per chapter:

• Tokens - what it puts in the context window, or spends.
• Turns - how many round trips before the job is done.
• Blast radius - what it can touch if it is wrong.
• Recurs? - does the fix hold next session, or do you pay for it again?

Most tool documentation answers the first two and goes quiet on the last two. The last two are where the money is.

The curriculum

Ten chapters, in reading order. Each one names the part of the gap it attacks, closes it, and hands you the bill.

1 · Getting started Install, authenticate, and trace one turn of the loop end to end - then ship a reviewed fix and watch the agent quietly break the money rule while doing it.

2 · Sessions & context The window is the only thing the agent knows, and it fills. Resume, fork, branch off, compact, and see exactly what you are throwing away each time.

3 · Approvals & sandbox You have to let it act before it fully knows. Codex splits that into two dials instead of one slider, so you can stop being interrupted without widening the blast radius.

4 · Rules (AGENTS.md) Written context is prepaid, and every line of it is rent you pay in every session. Work out what earns a line, then stop re-explaining the same three facts.

5 · Models & effort Run one task at every reasoning level and watch where deeper thinking stops buying anything. Ends in a table that says which dial for which shape of job.

6 · Subagents The search pollutes the window that has to hold the answer. Push it into an isolated context, fan it across worktrees, and pay the coordination cost honestly.

7 · Skills A procedure is knowledge the agent cannot rediscover from the code. Package the bank-import routine as a SKILL.md it invokes for itself.

8 · Extending Codex Two arguments: reach the agent does not have, over MCP, and the fact that nothing else in the loop can actually say no - which is what hooks are for.

9 · Automation The same loop with the approve beat removed. Run headless, wire it into CI, and learn which stage of an unattended run you cannot walk back.

10 · Daily workflow Stack the week's moves against a stated target, add up what each one saved, and see the whole ledger replayed in one table.

How to read it

The chapters form a chain. Each one opens on the problem the last one left you holding and closes by handing its own cost forward, so reading in order is worth real money in comprehension. Drop into a single lesson from a search and it still works; you will just miss the argument it sits inside.

There is no template stamped on the lessons. Some need a long worked trace, some need one command and a warning, some need a digression into why the tool behaves the way it does. The shape follows the material.

If you want the spec-level view of each primitive rather than the muscle memory, that is Foundations. If you use Anthropic’s CLI instead, the Claude Code course is this one’s sibling: same spine, different tool.

Start where the gap is widest, which is at the beginning. Chapter 1: your first hour with Codex.