How to research in ChatGPT and build in Claude Code

You can research a technical issue in ChatGPT and build in Claude Code by converting the result into a selected engineering briefing. RelayLink does not transfer the research session, hidden model context, or repository. It carries text only after you review the evidence and explicitly confirm the package.

Claude Code still needs separate access to the correct repository and branch. The handoff should look like a testable engineering task, not a transcript with “please implement” appended.

End research with a decision record

Before opening the coding phase, reduce the research to claims that can be checked:

• the observed problem;
• the likely cause and competing explanations;
• source names and URLs;
• quoted requirements or version details where precision matters;
• findings you personally verified;
• model inferences that remain uncertain;
• evidence that would disprove the recommendation.

RelayLink does not preserve ChatGPT's tabs, browsing results, or citation interface. It does not verify a source. If a page matters, include its name, URL, relevant section, and date as text. Claude Code may still need separate network permission to inspect it.

Write the implementation contract

Translate the findings into fields Claude Code can act on:

1. Problem: Describe the current behavior and impact.
2. Sourced findings: Tie each constraint to a source or label it unverified.
3. Requirements: List observable outcomes.
4. Non-goals: Exclude tempting adjacent changes.
5. Assumptions: Separate your statements from ChatGPT's inferences.
6. Repository state: Name the repository, branch, and commit.
7. Relevant paths: Point to likely code, tests, configuration, or docs.
8. Test target: Give the command and expected result.
9. First ask: State whether Claude Code should inspect, plan, implement, or diagnose.

A compact example might say:

• Problem: requests retry after a validation refusal.
• Finding: the provider documentation says only transport failures are retryable; URL and section included.
• Requirement: validation refusals return once without a second attempt.
• Non-goal: do not replace the shared HTTP stack.
• State: branch fix/retry-policy, commit abc1234.
• Paths: src/Transport/RetryPolicy.cs, tests/RetryPolicyTests.cs.
• Test: dotnet test tests/Transport.Tests.
• Ask: verify the finding against the current implementation, then add the smallest covered change.

These names are illustrative. Use the exact state of your own project.

Send the contract through your own inbox

Connect ChatGPT and Claude Code to the same RelayLink account. The account represents your email identity, so both clients can use its correspondence inbox.

From ChatGPT, compose a package addressed to your own account. Review every source claim, assumption, requirement, path, and command. Add an approved note that says what authority Claude Code has to change the project, then explicitly confirm.

The shorthand form opens a draft and never sends. After confirmation, the self-handoff appears unread until Claude Code pulls it. No contact request, second account, notification email, or magic link is involved. Your account page provides the portal fallback.

If Claude Code is not connected yet, use the current setup guidance in adding RelayLink to Claude or ChatGPT or check the running host's reference page.

Open the same repository separately

RelayLink does not upload source, attach files, clone a repository, check out a branch, or reproduce tool state. Claude Code must run in an environment that already has authorized access to the named repository and branch.

The package's path and commit are coordinates, not transported artifacts. Before editing, Claude Code should:

• confirm the working directory and branch;
• compare the current commit with the handoff;
• inspect the named paths;
• run the baseline test command;
• report mismatches or missing access.

If the branch moved after the research checkpoint, decide whether to update the brief or rebase the task. Do not silently implement against a different state.

Make uncertainty part of the test plan

Research uncertainty should shape implementation. If a provider behavior is undocumented, prefer a focused probe or characterization test before a broad change. If two sources conflict, ask Claude Code to preserve the current behavior until the conflict is resolved.

The code, tests, and authoritative documentation remain the sources of truth. The RelayLink briefing explains why the task exists and what you approved.

Check the provider boundary

Repository details, internal URLs, logs, and source excerpts may be confidential. Confirm that both AI providers and the Claude Code environment are allowed to receive them. RelayLink's review step does not replace employer policy, client authorization, or repository access controls.

For the broader workflow behind this technical template, read using ChatGPT for research and Claude for coding.