You can switch AI tools without copying the whole chat by sending a minimum sufficient briefing. RelayLink does not synchronize or import the conversation or hidden model context; it carries selected text only after you review it and explicitly confirm the handoff.
That matters when the current chat contains client names, rough ideas, abandoned prompts, and pages unrelated to the next task. Copying everything would expose more material and make the useful part harder to find. A minimal handoff keeps enough context to act correctly and nothing merely because it appeared in the same conversation.
A whole chat carries accidental material
Working conversations are shaped for exploration. They contain corrections, repeated drafts, speculative examples, and questions you decided did not matter. They may also contain data relevant to one phase but inappropriate for another provider.
A destination assistant has to infer which lines remain authoritative. An early proposal and its later rejection often look equally prominent in an export. The actual question can disappear beneath the history.
Copying less is not only about privacy. It improves clarity by forcing each included item to answer a question: what does the next assistant need in order to do the assigned work without guessing?
Build the minimum sufficient handoff
“Minimal” does not mean vague. A two-sentence prompt that omits a critical constraint simply transfers the cost into mistaken work.
Keep:
- the specific ask and expected response;
- cold-reader context;
- facts and where they came from;
- constraints and non-goals;
- assumptions, labeled as stated or inferred;
- accepted decisions;
- rejected paths whose return would waste time;
- uncertainty that should limit the conclusion;
- artifact locations and current test or review state.
Remove:
- social exchanges and prompt repair;
- duplicate drafts;
- tangents unrelated to the next ask;
- speculative content you rejected;
- confidential fields that the destination does not need;
- hidden shorthand you cannot explain.
If removing a section would make the second assistant guess, replace it with one direct sentence rather than restoring ten messages.
Review for both disclosure and meaning
Read the handoff twice. On the first pass, ask whether every person, organization, identifier, and excerpt is necessary and permitted for the destination provider. On the second, read as a cold recipient and ask whether the decisions and uncertainty are still understandable.
Check employer policy, contractual duties, and provider data settings before moving confidential work from one AI service to another. Human approval in RelayLink confirms that you chose the package. It does not grant a provider permission to process material that your rules prohibit.
The exact transcript still has a legitimate use when the conversation itself is evidence. If you need to show precisely what a model said, preserve the record through the appropriate sharing route. A briefing and a shared chat link solve different problems.
Send selected context through RelayLink
Connect ChatGPT and Claude to the same RelayLink account. The account is your email identity, not one model. Ask the first assistant to compose a package addressed to your own account.
Inspect the rendered ask, context, assumptions, options, decisions, open questions, and note. Remove accidental material, correct model inferences, and explicitly confirm. The shorthand only opens the draft. It never sends without that confirmation.
The package appears as unread inbound work until the other connected assistant pulls it. Since you are sending to yourself, there is no contact request, second account, email notification, or magic link. The account portal is the inbox fallback.
RelayLink sends the reviewed text only. It does not copy hidden context, memory, files, uploads, browser tabs, project spaces, repositories, branches, or tool state.
Reconnect only the artifacts required
Provide the destination assistant with separate authorized access to the artifacts it actually needs. For code, include repository, branch, commit, paths, and test command as text in the brief. For research, carry source names, URLs, dates, and uncertainty. RelayLink neither uploads those artifacts nor verifies the sources.
A link is a reference, not an instruction for the assistant to fetch it automatically. A path does not confer repository access.
Prefer a checkpoint over continuous mirroring
Do not try to keep two live chats synchronized message by message. Each assistant will develop its own working state, and repeated copying recreates the privacy and clarity problem.
Switch at a checkpoint, hand over the state that survived, and let one assistant own the following phase. The project continuity checklist shows how to define that checkpoint.
The result is deliberately smaller than the chat. It is not a lesser copy. It is a document written for the next decision.