If Your Account Disappeared Tomorrow, What Would Be Left?
This Part
A friend of mine ran all of his work through AI. His own projects, and also the work of his company, which was a dental company. Advertising copy, internal guidelines, draft contracts, plans and ideas he had spent months on, all of it was inside one account. One day the account was closed. Why it was closed is not our subject. The subject is that when that account closed, almost all of that work went with it, because none of it existed anywhere else.
This is not a distant event for anyone who works with AI even a little. Nobody decides to put all their work in one place, but gradually that is what happens. You build a project, you keep the prompt that worked right there, you add reference files to the project’s knowledge, and after a few months an important part of your work exists in only one place: an account you do not fully control.
Why Is Having One Place Dangerous?
Think of a practice whose patient records are on a single computer with no other copy. As long as that computer is healthy there is no problem, but when one hard drive fails, the entire treatment history of the patients is gone. Such a place is called a single point of failure: a place where, if that one thing fails, everything that depends on it fails too.
Your AI account is in the same position. It does not have to be closed; it can be hacked, its subscription can lapse over a payment problem, or you can delete a project by mistake. In none of these cases do you need to have done anything wrong. It is enough for everything to have only one copy.
What Should You Keep a Copy Of?
You do not need to copy every conversation. The real value is not in the conversations, it is in what you built inside them: the prompts that reached their final version, the workflows that worked, the decisions you made and the reasons for them, and the style rules that settled in for your texts. If it took three weeks and two hundred messages to build a prompt for validating articles, what you need is not those two hundred messages, it is the final version of the prompt and a few lines on why it ended up this way.
The best form for keeping these is a Markdown file, the same simple text file we saw in Chapter 6. You can give this file to any other model and continue from there, even if it belongs to another company. Making it is also the model’s job: in each project, ask it to produce a Markdown file from the instructions, prompts and decisions. Just know that the model does not see your whole account at once and only has access to the project it is in, so this is a habit you carry out yourself every so often, not an automatic service.
But where do you put this file? If it stays inside the same account, it is lost when the account is lost. It has to live outside the account. My suggestion is GitHub.
What Is GitHub?
GitHub is an online space for keeping files. It was built so programmers could keep their code there, but it works just as well for any text file, including these Markdown files. Files on GitHub are kept in what is called a repository, or repo for short. Think of a repo as an online folder where your backup files collect.
The first question is usually whether your information becomes public. No. If you create the repo as private, only you can see it, and creating a private repo on GitHub is free. Even so, the rule of Chapter 7 holds here too: you do not put patient information, passwords or access keys anywhere, even in a private repo.
What sets GitHub apart from an ordinary folder or a flash drive is that it keeps not only the latest version of a file but the earlier versions as well. If you change the prompt you wrote to explain post-implant care and the result gets worse, or you delete part of your instructions by mistake, the earlier version is still there and you can go back to it.
You Do Not Need to Know GitHub
GitHub may look hard at first, because its environment is designed for programmers. But you are not going to work with it yourself, because you are working with AI and AI knows how to do this. In Claude, this is done by Claude Code, and in ChatGPT, by Codex. Both can write directly to GitHub. All you have to do is tell them to make a Markdown file from this project’s instructions, prompts and decisions and put it in your repo on GitHub. Making the file, putting it in the repo and keeping the earlier versions are all done by the tool itself.
Why Is GitHub Alone Not Enough?
GitHub is also, in the end, an account, and you do not own it. If all your backup copies are only there, the first problem repeats itself and this time GitHub is your single point of failure. That is why you also need a copy on your own laptop. You can download the files from GitHub and bring them to the laptop.
This arrangement is very convenient in practice. In the middle of work, even from your phone, you tell the AI to make the file and put it on GitHub. It takes a few seconds, and anything that has just reached a result is recorded at that moment. Then, at your leisure, when you are sitting at your laptop, you move the files from GitHub to the laptop. If you had to save directly to the laptop every time, you would often put it off because you are not at the laptop, and a lot would be left behind. With this method, your work always exists in at least two separate places: one on GitHub and one on your own laptop.
The next part explains step by step how to create a repo, connect the AI to it, and bring the files from GitHub onto your laptop.