Field note
Move Chat Between AI Assistants Safely
Treat a copied conversation as a new data transfer with its own trust boundary, and carry over a short verified brief rather than the whole thread.
Treat a copied conversation as a new data transfer with its own trust boundary, and carry over a short verified brief rather than the whole thread.
Why people move chats, and what goes wrong
The usual reasons are practical. ChatGPT has hit a usage limit and the deadline is today. A Claude conversation has grown too long and the model has started losing the thread. The business has chosen a different vendor. Or one assistant is better at the next step, such as spreadsheet work or code.
The move itself is trivial: select all, copy, paste. That is the problem. A long thread contains things nobody meant to send anywhere: a customer's name and address pasted in for a letter, an API key dropped in while debugging, a colleague's salary from an HR question, and half a dozen instructions the model was given early on that no longer apply. It also contains claims the first assistant made that were never checked, now dressed as facts. Paste it into a second tool and you have made a second copy of the personal data, given a second vendor the secret, and told the new model to trust the old model's mistakes.
Know what each tool keeps
Before moving anything, know what the source and destination retain, because the copy will live under the destination's rules. Claude lets account holders export their data from settings, which gives you a record of what was in the conversations you are leaving. Google publishes a Gemini Apps privacy hub that explains what activity is stored, for how long, and how to turn it off. ChatGPT offers a similar data export and a memory feature that can carry details between chats. If the destination has memory turned on, whatever you paste may persist beyond this one conversation.
On the business side, the rule is simpler: use the company's business plan on both ends. Moving a thread from a work account into a personal account, or the reverse, breaks whatever retention and processing terms the business relies on. Our note on ChatGPT memory and business privacy covers the settings to check.
Redact before you copy
Do not move the raw thread. Read it and produce a handover brief in a plain text document first. Strip:
- names, addresses, account numbers and anything else that identifies a customer, employee or supplier, replacing them with placeholders such as
[CUSTOMER]; - passwords, API keys, tokens, connection strings and internal URLs;
- pasted documents, contracts and emails that were only needed for one earlier step;
- anything you would not put in an email to the vendor's support desk.
If a secret was already pasted into the first chat, rotate it. Moving the chat does not undo the disclosure, and neither does deleting the thread.
Write a brief, not a transcript
The destination assistant does not need the journey; it needs the current state. A good handover brief fits on one screen and has five parts:
- Goal. One or two sentences on what you are trying to produce and for whom.
- Decisions made. The choices you have settled on, stated as facts you own: audience, format, tone, constraints, deadline.
- Verified inputs. Figures, quotes and sources you have checked yourself, with the source link where you have one.
- Unverified claims. Anything the first assistant asserted that you have not checked, labelled as such. This is the part people skip, and it is the part that stops errors compounding.
- Current draft. The latest version of the output, and nothing earlier.
Leave out the first assistant's instructions to itself, its apologies and its reasoning. Leave out your own early prompts that were superseded. The brief should read the same whether the next reader is a model or a colleague picking up the job.
Check the answer in the new tool
Different assistants make different mistakes, and the second one will take your brief at face value. So in the new conversation, before continuing, ask it to list the assumptions it is working from and compare that list with your brief. Then re-verify the two or three facts that matter most in the output, from the original source, not from either assistant.
If the task involves code, run it. If it involves numbers, recompute one of them by hand. If it involves a customer communication, read it as the customer would. The ChatGPT, Claude and Gemini comparison note describes where each tool tends to be strong, which helps you decide what to check.
What to write into policy
For a small business, three lines in the AI usage guidance are enough:
- Conversations are moved between assistants only as a redacted handover brief, never as a pasted transcript.
- Any secret that has been pasted into an assistant is treated as disclosed and rotated.
- Work conversations stay in company accounts; personal accounts are not used to continue work started at work, and vice versa.
Add one operational habit: when a thread is getting long enough that you are thinking of moving it, write the brief anyway. Starting a fresh conversation with a clean brief in the same tool usually fixes the "it has lost the plot" problem without changing vendor at all.