Two AIs Are Better Than One: Meet AI Relay
What's happening
A research paper posted to arXiv in October put a simple question to the test: when one AI writes something, does it help to have a different AI review it?
The setup was 30 documents with 150 planted errors and 900 review sessions. The finding that matters for the rest of us: different models catch different mistakes. One review by the same model plus one review by a different model caught 56.7% of the planted errors. Two reviews by the same model caught 42.7%.
One catch worth knowing. A lightweight, budget model did not help as the second reviewer. The second set of eyes has to be a strong one. (It is also a preprint, so treat it as early evidence, not the final word.)
What I built: AI Relay
I already had a habit of doing this by hand. Ask Claude, copy the answer into ChatGPT, ask it to poke holes, copy that back. It works. It is also tedious.
So I built a desktop app with Claude Code that does the copying for me. I call it AI Relay.
What it does: It opens two AI chat windows side by side (any two of Claude, ChatGPT, and Grok), takes the newest answer from one, wraps it in instructions, and hands it to the other. Then it brings the response back. Round after round, the two AIs work the problem until the answer gets better.
The main features:
Six relay styles - Peer Review, Debate, Improve, Collaborate, Fact Check, and Custom. Each one tells the receiving AI what job to do with the other's answer.
You choose the direction and the rounds - one way, or back and forth, for 1 to 10 rounds.
Supervised mode on by default - before anything is sent, you see a preview and can Send, Edit, Skip, or Stop. Automatic mode is there when you trust the setup.
Safety stops - a Stop button, an Emergency Stop (with a keyboard shortcut), and automatic pauses for repeated answers, slow responses, oversized responses, or a login page.
A transcript of every session - saved on your own computer so you can see exactly what each AI said.
One sign-in per AI - you log in to each site yourself, once. The app never sees your passwords.
A real example: I asked Grok how often a small business should back up its files. It gave me one sentence: back up at least daily. I ran that through AI Relay to Claude with the Improve style. Claude came back with a real plan - backup frequency by type of data, the 3-2-1 rule, a reminder to actually test restoring a backup, and a catch I see people miss all the time: Google Drive and OneDrive sync is not a true backup. If you delete a file, sync deletes it everywhere.
Then Claude flagged one of its own numbers as a figure of speech and told me to remove it before publishing. That is the kind of thing that happens when AIs are asked to check work instead of just produce it.
How it helps you:
Better answers on anything that matters - a policy, a proposal, a process write-up, a tricky email.
Fewer confident-sounding mistakes making it into your final work.
No more copy-paste between tabs.
A record of how the answer got where it ended up.
A quick caution: AI Relay works by driving the regular chat websites in a browser window. Check each AI provider's terms of use before you run something like this, keep it to your own personal use, and leave supervised mode on.
Try this prompt
You do not need to be a developer to build this. Paste the prompt below into Claude Code and let it work. Plan on a few sessions, not a few minutes.
I want you to build a desktop app called AI Relay for my computer. Check whether I am on Mac or Windows and build for that. The app passes answers back and forth between two AI chat websites so they can review and improve each other's work, without me copying and pasting. I am not a developer, so explain each setup step in plain language. Use Electron and TypeScript for the app and Playwright to control the browser windows. Core features: 1. Let me pick any two of Claude, ChatGPT, and Grok. Open each one in its own browser window with a saved profile so I only sign in once. I type my own passwords into the real sites. The app must never see or store them. 2. Wait until an AI has finished writing, grab its newest answer, wrap it in instructions, and paste it into the other AI. 3. Include relay styles I can pick and edit: Peer Review, Debate, Improve, Collaborate, Fact Check, and Custom. 4. Let me choose the direction (A to B, B to A, or back and forth) and the number of rounds (1 to 10, default 3). One round is one full back-and-forth. 5. Supervised mode is on by default: before each send, show me a preview with Send, Edit, Skip, and Stop buttons. Automatic mode is an opt-in checkbox. Safety: - A Stop button and an Emergency Stop button (plus a keyboard shortcut). Both halt the relay but leave the browser windows open. - Pause if the same answer repeats, if an AI takes longer than 5 minutes, or if an answer is longer than 30,000 characters. - If a site shows a login page or CAPTCHA, or you cannot find the message box, pause and tell me what happened in plain English. Records: - Save a readable transcript of every session to a folder on my computer, outside the app itself. How to build it: - Inspect the real Claude, ChatGPT, and Grok pages to find the right text boxes and buttons. Do not guess. - Build one piece at a time. After each piece, run it and test it against the live sites before moving on. Do not mark anything done just because the code is written. - At the end, package it as an app I can double-click, and give me a short list of anything you could not test.
The AIM take
The research and my own testing land in the same place: a second AI is worth it, as long as it is a strong one and a person stays in the loop.
What surprised me most was the build itself. Claude Code wrote the app, but testing is what made it work. Running it against the live sites turned up six real bugs that would never have shown up just reading the code. That is why the prompt above insists on testing each piece before moving on. Same lesson as the paper, really - the check matters as much as the work.
You do not need the app to start. Next time an AI hands you something important, paste it into a different AI and ask it to find the weak spots. See what it catches.
If you want help figuring out where a second set of AI eyes would save your team time, book a free workflow review at intellmanage.com.




Comments