You can continue a ChatGPT project in Claude, but RelayLink cannot import the project or preserve its chat state. It carries a selected briefing as text only after you review and explicitly confirm it; files, model memory, browser tabs, repositories, branches, and tool settings stay outside the handoff.
Treat the move as a controlled migration. Freeze the artifacts and decisions that matter, arrange Claude's access to them separately, then define a clean first task.
Freeze a checkpoint
Choose a stopping point before you start moving material. Finish or explicitly pause the current task, save the artifacts, and record the state you are willing to call the baseline.
For a software project, that baseline might be a commit with a known test result. For a writing project, it might be a versioned draft with approved sections and comments left on the unsettled ones. For research, it might be a source list and a recommendation marked provisional.
Do not keep making decisions in ChatGPT while Claude starts from an older snapshot. If both assistants are active during the move, name which one owns each open item.
Inventory artifacts separately from conversation
List what the project actually depends on:
- repository, branch, and commit;
- documents and their versions;
- datasets or source material;
- requirements and acceptance criteria;
- test commands and current results;
- external systems or tools required for the work.
These artifacts are not the ChatGPT conversation. They need their own authorized storage and access path. RelayLink does not upload files, copy a project workspace, move a repository, or reproduce tool settings. Claude must receive separate access to anything it needs to inspect.
The conversation can still matter when an exact exchange is evidence, such as a reproduction or an instruction you must retain verbatim. Keep that record where it belongs, but do not force Claude to mine every chat for project state.
Write the migration brief
Compose a cold-reader explanation of the checkpoint:
- Purpose: what the project is for and who the result serves.
- Current state: what is complete, in progress, and blocked.
- Accepted decisions: what should remain unchanged unless named evidence reopens it.
- Rejected paths: what was tried or considered, with reasons.
- Constraints: policy, technical, time, format, and scope boundaries.
- Assumptions: what you stated versus what ChatGPT inferred.
- Uncertainty: facts or sources that still need checking.
- First Claude task: one bounded output and its acceptance test.
Review every model-written claim. A migration brief that confidently preserves a mistaken inference makes the mistake harder to detect, not easier. For research claims, cite source names and URLs as text, preserve uncertainty, and verify them independently; RelayLink does not verify sources.
Transferring a ChatGPT research session to Claude gives extra detail for sources and uncertainty.
Hand the brief to your own account
Connect ChatGPT and Claude to the same RelayLink account. From ChatGPT, draft a package to your own email identity. Inspect the package, edit anything that does not represent your position, and explicitly confirm it.
The shorthand opens the draft and never sends. After confirmation, the package appears as unread inbound work in your RelayLink inbox until Claude pulls it. A self-handoff needs no contact request or second account, and it sends no notification email or magic link. The portal at your account remains the fallback.
RelayLink carries the reviewed briefing, not the ChatGPT project. That boundary prevents the convenience claim from outrunning the product.
Start Claude with verification
Claude's first action should orient against the real artifacts, not immediately extend them. Ask it to confirm the named version, run or inspect the stated test, and identify any mismatch between the brief and the artifact.
Then give it the bounded first task from the migration brief. This catches stale paths, missing access, and unsupported assumptions before they become more work.
If Claude cannot access a source or repository, solve that through the artifact's own permission system. A URL or path in RelayLink is text, not authorization and not an instruction to fetch.
Check policy before changing providers
A tool migration may also be a data transfer between AI providers. Check employer rules, client obligations, and provider settings before including confidential material. RelayLink's human approval confirms what you chose to send; it does not grant either provider permission to receive it.
For a vendor-neutral version of the same checkpoint, use continuing a project in a different AI assistant. The core rule stays the same: move a reviewed state of play and the artifacts it references, not the illusion of a portable session.