Your agent branches, rebases, merges and force-pushes on your behalf. Readback teaches you the model underneath, so you can read what it actually did — and know whether it was right.
In air traffic control, a readback is the procedure where the receiving station repeats a message back to the transmitting station, to confirm it was received correctly. The controller issues a clearance; the pilot repeats it back; the controller listens to that repetition and corrects anything wrong. It is mandatory for anything safety-critical — runway crossings, altitudes, headings.
A readback does not only confirm the pilot heard correctly. It also catches the case where the controller said something they did not mean. It traps errors in both directions at once.
That is exactly the situation you are in. When you read back what your agent did, you catch two different failures with one action: what you misunderstood, and what it got wrong.
The word has a second technical home, which this audience tends to know already. In hardware, readback verification is reading a value back off a register after writing it, to confirm the write actually took. Same idea, same word.
This is not a course that teaches you to type commands. You are not going to type them — your agent will, and that is a reasonable way to work. This treats the delegation as permanent, rather than as a bad habit to be corrected.
What it teaches is the model underneath: what a commit actually is, what a branch actually is, what a rebase actually does to your history. Enough that you can open a repository an agent touched and say, without hedging, exactly what happened to it.
Every lesson is a sequence of steps, and every step after the first ends in something you can verify yourself. Every command shown was executed before it shipped, and its output pasted back verbatim. Nothing here is a plausible-looking example.
Git is the first course, not the whole product. The same problem — competent people who have become passengers on tools an agent drives — shows up across the toolchain, and the discipline is the same one every time. Read back what happened. Check it against the model. Then trust it.