To continue a project in a different AI assistant, stop at a clean checkpoint and hand over a selected briefing that you review and explicitly confirm. RelayLink carries that text; it does not synchronize or import the first assistant's chat, hidden context, or model memory.
The direction can be ChatGPT to Claude, Claude to ChatGPT, or another permitted pair. Availability, workflow, policy, or preference may drive the switch. Continuity comes from the checkpoint, not from making two assistants pretend they share one session.
Choose a checkpoint that can stand alone
Switch after a meaningful unit of work, not halfway through an unexplained correction. A good checkpoint has four properties:
- the artifact is saved and identifiable;
- the current task is finished or clearly paused;
- accepted decisions are written down;
- the following task can be stated without “as discussed above.”
For code, use a commit or an intentionally named working-tree state with a known test result. For a document, use a version with approved and unsettled sections marked. For analysis, freeze the evidence set and say which conclusion remains provisional.
If you must switch during a failure, make the failure the checkpoint: exact command, actual output, expected result, paths involved, and attempts already ruled out.
Describe state, not conversation order
The destination assistant does not need a replay of how you reached the checkpoint. It needs a map of the state:
- Project purpose and intended audience.
- Artifact locations and exact versions.
- Requirements, constraints, and non-goals.
- Decisions already accepted.
- Rejected approaches and their reasons.
- Assumptions labeled as stated or inferred.
- Open questions and current uncertainty.
- One next task with a completion test.
This structure works in either direction because it does not depend on Claude or ChatGPT interpreting the other model's conversational habits. It is written for a cold reader.
A briefing provides the selected ask, context, assumptions, options, decisions, questions, and approved note. It is the correspondence layer, not the project itself.
Hand off through your RelayLink inbox
One RelayLink account can connect to both assistants because it represents your email identity. From the assistant holding the current session, compose a package addressed to your own email.
Review the rendered briefing carefully. Confirm that inferred assumptions are labeled, rejected paths have reasons, and the next task is narrow. Then explicitly confirm the package. The shorthand form creates a draft and never sends it.
The self-addressed package appears unread in the account inbox until the destination assistant pulls it. It produces no contact request, second account, notification email, or magic link. You can use the account portal if you need to inspect the package without either assistant.
If the destination is another person rather than another one of your assistants, this self-handoff pattern does not apply. Their account follows the ordinary contact and consent rules.
Reconnect the artifacts independently
RelayLink carries the briefing as text. It does not carry a repository, branch, file, attachment, browser tab, memory, project space, or tool configuration.
Give the destination assistant separate, authorized access to the same artifacts. Put repository and branch names, commit identifiers, file paths, document versions, source URLs, and test commands in the handoff so it knows what to open after access is established.
Do not treat a path as an upload or a URL as an automatic fetch instruction. RelayLink does not inspect those targets, verify research sources, or preserve the first assistant's browsing state.
Begin with an orientation pass
Ask the destination assistant to verify the checkpoint before changing anything. It should confirm the artifact version, restate the fixed constraints, run or inspect the baseline test, and identify mismatches between the brief and the actual project.
Only then should it begin the stated task. This short orientation is cheaper than discovering later that the second assistant opened the wrong branch or treated an inference as a requirement.
Keep the checkpoint portable and permitted
Write project facts in ordinary language rather than commands tied to one model. That makes the same checkpoint useful if you switch again or continue without an assistant.
Portability does not erase provider boundaries. Before moving confidential material between AI services, check employer policy, client obligations, and provider terms. RelayLink records your review of the briefing; it does not authorize either provider to process data it should not receive.
For direction-specific migration details, use moving a ChatGPT project to Claude or moving a Claude project to ChatGPT. The checkpoint stays the durable center in both cases.