Switching AI tools without starting over

Switching tools without starting over

At some point you are going to change AI tools. Your department buys a license for something and IT rolls it out, so the account you spent six months getting dialed in is no longer the one you are supposed to work in. Or you hear another tool handles something better and you want to test it, but the thought of explaining your program from scratch again is enough to keep you exactly where you are.

That second one is the more common problem, because believing a switch means starting over keeps you on whatever tool you happened to try first. The good news is that this takes about ten minutes to export your data and get the new tool set up.

What you are actually moving

Two different things live in these tools, the first is everything you wrote down on purpose. Project instructions, standing instructions, any documents you saved. That is text you own, so it copies anywhere and there is no trick to it.

The second is what the tool figured out about you on its own, which is the memory we covered a couple issues back. That is the part people worry about and the part that actually takes a few steps to move. It is also worth saying that a fair amount of what is stored in there is probably outdated or not that useful, so this is a good excuse to look at it and decide what deserves to come with you.

Inside the Lab: the move

Step one, get the old tool to write out what it knows. This works in any of them. Open a chat in whatever tool you are leaving and paste this:

"I'm moving to a different AI tool and I want to take my context with me. Don't limit this to your saved memory. Look back through our conversation history and pull out the patterns too, including things you picked up on that were never explicitly saved. Cover all of this:

• What you have stored about me, in your own saved wording, with dates if you have them.
• Who I am and what I do. My role, my setting, who I work with, anything about my background that keeps coming up.
• What I'm working on. One entry per project, with what it's for and where it currently stands.
• How I want you to work with me. Tone, format, length, any standing rules I've given you about what to always do or never do.
• What you've figured out about how I think. How I make decisions, what I tend to push back on, what I care about, how I like to work through a problem.
• What I keep coming back to. The topics and questions that show up over and over.

Put all of it in a pdf file so I can download it. Label each item as either something you saved or something you inferred, because I want to know the difference. After you generate the pdf file, tell me what you pulled from, what you think is missing, and which parts you're least confident about."

You can also look at the stored memory directly before you run the prompt. In ChatGPT it is Settings, then Memory, where there is a memory summary button. In Claude it is Settings, then Capabilities, then view and edit your memory. Not every tool makes that easy to find, which is another reason the prompt above is the more reliable way to see what is in there.

Step two, read it and cut. This is the step most people skip. Anything about a job you no longer have, a team you no longer coach, or a system you stopped using should not come with you, because stale context follows you into the new tool and causes the same problems it was causing in the old one. Take out anything sensitive while you are in there too.

Step three, load it into the new tool. Worth knowing that none of these tools have a one click export, which is why step one exists. Getting your context out always means asking the tool to write it out for you.

Getting it in is easier. Claude and Gemini both have an import memory feature that walks you through it, and each supplies its own version of the extraction prompt to run in whatever tool you are leaving. Check settings or the home screen in whichever one you are moving to. Anthropic labels theirs experimental and says it may not pick up everything, and that it prioritizes work related context, so plan to verify it rather than assume it worked.

If the tool you are moving to does not have an import option, you do it by hand and it takes about a minute. Paste your edited context into the custom instructions, start a project and put it in the project instructions, or open a chat and paste it as your first message with a line saying this is background on how you work.

Step four, move the things you built. Any project instructions, saved documents, or standing prompts get copied over the same way. These are just files and text, so it is mechanical.

Step five, check that it took. Open a fresh chat in the new tool and ask it what it knows about you. If something important is missing, add it manually rather than assuming it will pick it up later.

Where the coach stays in charge

The reason this works at all is that you did the writing somewhere along the way. If you never wrote standing instructions and never set up a project, then there is genuinely nothing to move and a switch does mean starting over.

That is the argument for doing the setup work even when it feels like overhead. You are not configuring somebody's product. You are writing down how you work, in a form you can hand to anything. That stays true no matter which company your athletic department signs a contract with next year, or for whatever new AI tool you want to try out next.

The Coach's Scouting Report

• Signal. Your context is not locked inside the tool. Getting it out takes one prompt, and the move takes about ten minutes in any direction.
• Noise. Staying with a tool because switching feels expensive. Dragging stale context into a new tool because you did not read it first.
• Next Rep. Run the extraction prompt in whatever tool you use most, even if you are not switching. Read what comes back, delete what is no longer true, and keep a copy of the rest somewhere outside the tool.

See you next Sunday.