Why split the work between two AIs
Different AIs are good at different parts of a job. A chat app is a comfortable place to explore options, ask "what about…" and settle on a plan. A coding agent like Claude Code works inside your project: it reads the files, runs the tests and makes the changes. Doing both in one place usually means compromising on one of them.
The hard part is the handoff. Pasting a long plan into a terminal loses structure, and when the build is done you have to explain the result again to get a review.
The loop
1. Plan in ChatGPT
Talk it through as usual. When you're happy with the plan, ask ChatGPT to save it as a handoff:
Save this plan to Relay for the project "Web store" and hand it to Claude Code.
Relay stores a structured brief: the goal, the decisions made, open questions and the next steps, and marks it as waiting for Claude Code.
2. Build in Claude Code
In your project folder, start Claude Code and say:
Pick up the latest plan from Relay and build it. Post a progress update when you're done.
Claude Code loads the brief (not a wall of chat history), builds it, and adds a progress update saying what it changed and what's left. The conversation moves to review.
3. Review back in ChatGPT
Review the latest progress in Relay for "Web store" and add your review.
ChatGPT reads the plan and the update together and records its review. If it asks for changes, the work goes back to building, and Claude Code picks up the review next time. When you're done, mark it Done in the dashboard.
Make Claude Code check Relay on its own
Add one line to your project's CLAUDE.md, and Claude Code looks for the latest plan at the start of every task:
At the start of a task, check Relay for the latest plan for this project. When you finish, add a progress update to Relay.
Claude Code also has Relay shortcuts: /mcp__relay__pickup loads what's waiting for it, and /mcp__relay__handoff saves the current work.
Set it up
- Sign in to Relay with Google.
- Connect ChatGPT: step-by-step.
- Connect Claude Code with one command: step-by-step.
The same loop works with other pairs: plan in Claude and build in Cursor, or plan in Gemini and build in Codex. See all supported tools.