You can give Claude the relevant result of work you started in ChatGPT, but the assistants cannot literally share context. There is no common context window to expose, and RelayLink does not synchronize model memory behind the scenes. It carries only selected text in a briefing that you review and explicitly confirm.
That composed account of the state that matters is different from the state itself. The distinction determines whether the next assistant receives a useful starting point or a pile of messages.
Why the context itself cannot cross
A chat transcript contains prompts and responses. Working context is broader. It includes the correction that changed every answer after it, an option you stopped discussing, a priority the model inferred from your objections, and shorthand that only makes sense because you remember an earlier meeting.
Some of that state was never written down. Some may have been compressed during a long session. Even a perfect transcript would still be indexed to what you know and what the first assistant could access.
Claude cannot import that state from ChatGPT, and ChatGPT cannot import it from Claude. Neither a share link nor RelayLink changes that. The full explanation of context as accumulated state is useful when an export keeps feeling as though it should be enough.
Turn the working state into a brief
Before switching assistants, ask what the next one must know to answer one defined question. Build the handoff around that answer:
- State the next ask in one sentence.
- Explain the situation for a reader who saw none of the first chat.
- List facts with their sources or locations.
- Label assumptions as stated by you or inferred by the assistant.
- Record decisions and the options you rejected.
- Name what remains uncertain.
- Add the wording you approve as your note.
This is composition, not automatic extraction. Review it for missing constraints and confident-sounding inferences. The first assistant is well placed to draft the brief because it has the working session, but you are the only person who can confirm that the result represents your position.
A RelayLink briefing gives those pieces a consistent shape without attaching the private conversation.
Share it through your own inbox
Both assistants may connect to the same RelayLink account because the account represents your email identity, not one model. Address the reviewed package to your own account. It appears inbound and unread until an assistant on the account pulls it.
The flow still has three distinct moments: draft, human review, explicit confirmation. The shorthand form opens the draft and stops there. RelayLink does not send automatically because the recipient happens to be you.
A self-handoff creates no contact request and sends no email notification or magic link. There is only one participant. If you need to inspect it without an assistant, use the inbox on your account page.
Be exact about what did not move
The package carries selected text. It does not carry either model's hidden instructions, memory, browsing session, uploads, files, tabs, repository, branch, or active tools. A source URL in the brief remains text; RelayLink does not verify it or tell the receiving assistant to fetch it.
For code, arrange repository and branch access separately. Include the relevant path, commit, current failure, and test command so the second assistant can orient itself after it opens the same working tree.
For confidential material, check whether both AI providers are permitted destinations under your employer's policy. RelayLink approval says that you approved the briefing. It does not override a data-handling restriction.
Test whether the handoff is complete
Read the brief as if the first conversation had disappeared. Could you identify the decision, distinguish evidence from guesses, and know what response is wanted? If not, the second assistant will compensate by guessing.
Do not solve that by adding the entire transcript. Add the missing fact or decision directly. The goal is not to make both assistants share a mind. It is to make the boundary explicit enough that a separate mind can continue the work.
For a broader division of roles around that boundary, use the workflow in how to use Claude and ChatGPT together.