You finish a research session in ChatGPT with twelve promising sources, three disputed claims, and a rough conclusion. You want Claude to write the piece. If you paste the entire chat, Claude has to discover which citations you checked, which claims you rejected, and which audience you actually mean. If you send only the conclusion, the evidence disappears.
The handoff between research and writing should be an evidence brief, not a conversation export.
Freeze the evidence before changing assistants
Research produces more than a topic. It produces a set of claims with different origins and different levels of support. Capture that structure while the source trail is still in front of you.
For each source you may rely on, record:
- Title, author or organization, publication date, and URL when known.
- The exact claim the source is being used to support.
- A short quotation when wording matters.
- Whether you opened and checked the source yourself.
- Any limitation, conflict, missing context, or reason for doubt.
“ChatGPT found this” is not a verification status. Neither is a plausible citation. If you did not open the source, write “not checked” rather than allowing the writing assistant to treat it as settled evidence. Preserving sources across assistants works because the ledger keeps origin and uncertainty beside the claim.
Turn findings into a writing brief
Claude also needs to know what the evidence is for. Add the editorial decisions that a source list cannot supply:
- Audience: who should understand or act after reading.
- Primary claim: the one sentence the draft should support.
- Scope: what the piece will and will not attempt to establish.
- Evidence map: which checked sources support, complicate, or contradict each major point.
- Uncertainty: what remains inferred, disputed, or unverified.
- Draft shape: the desired form, length range, voice, and next ask.
Keep human-confirmed facts separate from AI-inferred claims. ChatGPT may synthesize a pattern that is useful but not directly stated by any source. Label it as an inference. Claude can then phrase it cautiously, seek another source, or leave it out.
Use a self-handoff as the checkpoint
Once the evidence brief is ready, draft it in ChatGPT to your own email on the account both assistants use. Review every check status and quotation, then explicitly confirm. Because sender and recipient are the same account, there is no second account, contact request, email, or magic link. The package remains unread until any assistant pulls it; the portal inbox is the fallback.
What crosses is only the approved structured brief. RelayLink does not copy the ChatGPT session, preserve browsing state, upload research files, or cause Claude to open any URL. A link in the package is a reference, not proof that either assistant visited or validated it. Why context does not transfer explains why composing this checkpoint is different from exporting a chat.
Ask Claude to draft from claims, not confidence
The writing ask should tell Claude how to handle evidence:
- Attribute claims where attribution matters.
- Do not strengthen tentative findings into facts.
- Surface meaningful counterevidence rather than smoothing it away.
- Mark gaps that require your decision or another source.
- Keep quotations tied to the recorded source.
Claude may still misread a source, invent connective tissue, or produce a citation that looks complete but is not. One model handing material to another does not create independent verification.
Check the draft against the ledger
Read the draft with the evidence brief beside it. Open the sources yourself. Confirm quotations, dates, authorship, and whether each central claim says no more than its support allows. Check that uncertainty survived the change in voice.
Do not include confidential notes, customer data, subscription-only text, or protected research material unless the provider and receiving environment are authorized for it. Prefer source references and short necessary quotations over copied documents.
Use this split when research has reached a reviewable evidence checkpoint and drafting is a separate job. Keep the work in one authorized environment when Claude cannot access a required source or the central evidence is still unresolved. This is not a claim that ChatGPT researches better or Claude writes better; the ledger keeps either workflow accountable to what you actually checked.