To transfer useful work from ChatGPT to Claude, close the research phase and hand Claude a selected briefing, not the session itself. RelayLink does not import the conversation or hidden model context; it carries text that you review and explicitly confirm.
Begin with the result, then expose enough evidence and uncertainty for Claude to evaluate the conclusion. A full session gives Claude the order in which the research happened rather than a clear account of what survived it.
Close the ChatGPT phase first
Do not hand off while the question is still changing every few messages. Ask ChatGPT to draft a checkpoint that answers:
- What decision or deliverable is this research meant to support?
- Which findings matter to that decision?
- What source supports each finding?
- Which statements came from you?
- Which assumptions did ChatGPT infer?
- What remains unknown or contested?
- Which paths were considered and rejected?
Review that checkpoint against the actual sources. Remove claims that only repeat a model's confidence. If a finding depends on an inaccessible page, a stale document, or a calculation you have not checked, say so directly.
This step does not make the research correct. It makes the boundary between evidence and interpretation visible.
Write for Claude as a cold reader
Claude has none of the ChatGPT session. Avoid references such as “the second option,” “the report above,” or “the concern we discussed.” Expand each one.
A strong handoff can follow this order:
- Objective: the decision or output you need.
- Current conclusion: what the research supports so far.
- Evidence: source names or URLs beside the claims they support.
- Assumptions: separated into stated and inferred items.
- Rejected paths: what was ruled out and why.
- Uncertainty: what could change the conclusion.
- Ask for Claude: review, choose, challenge, or produce a specified artifact.
Make the ask narrow enough to judge. “Continue the research” invites another broad session. “Challenge the recommendation, identify unsupported claims, and return a decision with the two strongest risks” defines an endpoint.
Send the reviewed briefing to yourself
Connect ChatGPT and Claude to the same RelayLink account. An account is your email identity, so it can serve both assistants without creating a second mailbox.
From ChatGPT, compose a package addressed to your own RelayLink email. Inspect the rendered briefing, correct the assumptions and source descriptions, then explicitly confirm it. The shorthand form only creates the draft. It never skips review or sends on its own.
The package appears as unread inbound work for your account until Claude or another connected assistant pulls it. A self-handoff sends no email notification, creates no magic link, and needs no contact request. The self-briefing guide explains that inbox behavior.
Preserve evidence without overstating it
RelayLink carries source names and URLs as text. It does not carry chat state, model memory, files, repositories, branches, browser tabs, retrieval results, or a citations panel. It does not verify that a URL supports the claim beside it. Claude may also lack access to the source.
Write “Source: regulator guidance, section 4, URL…” rather than “verified fact” unless you performed the verification. Preserve qualifications such as sample size, date range, jurisdiction, and whether a statement is a direct quote or your interpretation.
Links in a package are references. They are not instructions for automatic fetching. If Claude needs the underlying documents, provide them through a separately authorized route and follow the relevant data rules.
Keep the human decision visible
The most important transfer is not the model's summary. It is your position on what should happen next. Use the approved note to state what you accept, what you do not accept yet, and what authority Claude has in the second phase.
If confidential work would move from one AI provider to another, check your employer's policy and both providers' terms first. Approving a RelayLink package does not grant permission that your organization withheld.
This pattern is especially useful when the second phase is implementation. Using ChatGPT for research and Claude for coding adds the requirements, non-goals, repository references, and test target that a coding handoff needs.