To transfer useful work from Claude to ChatGPT, give ChatGPT separate access to the current artifact and a selected briefing about its state. RelayLink does not copy the Claude session or hidden model context; it carries text only after you review it and explicitly confirm the handoff.
ChatGPT does not need every proposed edit. It needs the constraints that shaped the artifact and a question it can answer without repeating the first session.
Start with the current state
Open the handoff with a concrete description of what exists:
- the deliverable and its purpose;
- what is complete, partial, or failing;
- the decisions already accepted;
- the artifact locations;
- the current test, review, or acceptance result.
For code, name the repository, branch, commit, relevant paths, test command, and observed failure as text. For long-form work, name the document version, section boundaries, intended audience, approved outline, and unresolved passages.
Avoid chronology unless it explains a constraint. “Claude tried three approaches” is less useful than “Approach B was rejected because it violates the offline requirement.”
Separate constraints from preferences
The second assistant must know which boundaries are fixed and which are open to challenge.
Label legal, security, compatibility, format, deadline, and scope limits as constraints when they truly cannot move. Put style preferences, tentative architecture choices, and model suggestions in a separate list. If Claude inferred a constraint from your feedback, mark it as inferred until you confirm it.
This prevents two common failures. ChatGPT will not spend time proposing a path that is impossible, and it will not mistake one assistant's earlier preference for your final instruction.
Include rejected paths with short reasons. A dead end without its reason tends to return. A ten-message transcript of the dead end makes the reader discover the reason by archaeology.
Give ChatGPT one defined job
Choose the response you want before composing the handoff:
- review the implementation against five acceptance criteria;
- find contradictions in a draft;
- compare two remaining approaches;
- diagnose one named failing test;
- produce a migration checklist from the accepted design.
Do not ask ChatGPT to “pick up where Claude left off” without further direction. There is no shared marker for where that is. State what is closed, what is open, and what output would make the phase complete.
If you are moving the whole project rather than assigning a bounded review, use the fuller Claude-to-ChatGPT migration checklist.
Create a self-handoff in RelayLink
Claude and ChatGPT can connect to one RelayLink account because that account belongs to your email identity. From Claude, compose a package to your own email. Review the selected context, correct any inferred assumptions, and explicitly confirm the send.
The @relaylink shorthand stops at the draft. RelayLink does not copy the open conversation or deliver without your confirmation.
The package enters the shared account inbox as unread. It stays waiting until ChatGPT or another connected assistant pulls it. There is no contact request, second account, notification email, or magic link for this one-person thread. You can inspect the same inbox through your account page.
One account for multiple assistants explains the shared correspondence boundary and separate credentials.
Move artifacts through their proper channel
RelayLink carries the briefing as text. It does not upload source files, documents, attachments, repository contents, branch state, or tool output. Arrange ChatGPT's access to the artifact independently and according to your security policy.
A path such as src/payments/RetryPolicy.cs only tells ChatGPT where to look after it has access. A commit identifier tells it which state you mean; RelayLink does not fetch that commit. A document link remains a reference, not an invitation to retrieve it automatically.
For source-based work, cite the source name and URL in the briefing, preserve uncertainty, and verify the claim separately. RelayLink does not verify a source or preserve Claude's browsing tabs or model memory.
Check employer and provider rules before exposing confidential work to a second AI service. Your approval confirms that the handoff says what you intend. It does not establish that the destination is authorized.
Let the brief replace the chat, not the artifact
The working conversation can stay private. The artifact remains the source of truth for the work itself. The RelayLink package explains how to approach that artifact and what you want back.
That three-part boundary keeps the handoff legible: artifact over its own authorized channel, reviewed context through RelayLink, and a specific task for ChatGPT. Switching without copying the whole chat shows why the smaller record is often clearer as well as more private.