Back up before you let the AI work

Updated on

Before letting an AI change your project, you need two safety nets: a version tool that photographs every step (one command to go back), and a copy of the folder somewhere other than your machine. Without those, a single bad move can cost you weeks of work, permanently and with no recourse.

Why this is the first move, not the last

Everyone installs the safety net after their first disaster. That is human, and it is exactly the wrong way round.

A coding AI changes dozens of files in seconds. Most of the time that is what you wanted. On the day it is not, there are only two possible worlds: the one where you go back with one command, and the one where you redo weeks of work.

The choice between those two worlds is made before the accident. Afterwards, it has already been made.

Net number one — the project camera

A version tool is a camera for your whole project. At every step that works, one photo. Each photo is complete: every file, exactly as it was at that moment.

What that changes in practice:

The AI broke something? One command, and you are back at the last photo. No rebuilding from memory, no roughly-right: the exact previous state, to the comma.

You want to know what it changed? The tool shows you only the lines that differ between two photos. No more re-reading the whole project to find the suspicious edit.

The most widespread one is called Git. It is free, it works everywhere, and you do not need to understand it deeply: three moves are enough — photograph, compare, go back.

Net number two — the copy elsewhere

The version tool lives on your machine. If the machine dies, gets stolen, or the whole folder gets deleted, the photos go with it.

Hence the second net: a copy of the project somewhere other than your place. An online storage space you already use will do; the services built to host Git-tracked projects do even better, because they receive the photos as you go.

The simple rule: what exists in only one place only half exists.

The how-to, in three requests

You do none of this yourself — it is perfect work to hand to your AI, in this order:

1. “Set up version tracking on this project, and take a first photo of the current state.”

2. “Show me the exact command to go back to the last photo, and explain when to use it.” Write that command down. It is your emergency button.

3. “From now on, take a photo after every finished, validated step, with one sentence saying what changed.” Put that instruction in its startup sheet, so it survives closing the window.

The test that proves the net holds

An untested net is a hypothesis, not a protection.

Do the drill once, ten minutes, on a calm day: ask the AI to deliberately change something visible, see the change, then use your go-back command. Check that everything returned exactly as before.

The day a real problem arrives, you will not be discovering the procedure in a panic: you will have done it once already, and it will be a non-event.

Frequently asked questions

What is a “version tool”, in plain words?
A camera for your project. At every validated step, you take a photo of all the files. If something breaks later, you go back to any photo, identically, with one command. The best-known one is called Git; it is free and installed in ten minutes.
Can I ask the AI to install it for me?
Yes, and that is exactly the right reflex: “set up version tracking on this project and take a first photo”. Then ask it to show you how to go back, and write that command down somewhere.
Is a copy on a USB stick enough?
Better than nothing, but a stick gets lost, breaks, and usually stays plugged into the same machine. An online space (any storage service you already use) also protects against theft, failure and one bad click.

Sources

  1. Git documentation — getting started with version control — accessed