How I Trained Claude Code to Use ChatGPT to Double-Check Itself
I Love Claude. But It Measures Once and Cuts Twice.
I love Claude and especially the Opus model, but it’s definitely a “measure once and cut twice” kind of guy. ChatGPT and Codex are much more of a “measure twice and cut once” system.
That’s why my most used prompt with Claude has been:
“Are you sure? Double check everything and don’t make assumptions.”
Lo and behold, it usually has to correct itself a good bit. Having reviewer agents and skill setups helped, but separate agents don’t compare to having another model verify the output.
The Fix: Make Claude Verify Through ChatGPT
I recently set up my Claude Cowork to use ChatGPT to double-check itself, and it’s made the workflow so much more reliable. Anyone can do this with the right prompt.
There are two approaches depending on how you work.
For Technical Users: Claude Code + Codex CLI
If you’re technical and using Claude Code (and hence MD files), it’s best to train Claude to use the Codex CLI and ask Codex to verify all the info in the file. This skips going through the UI entirely.
You edit your CLAUDE.md (the instructions file Claude reads at the start of every session) to include a verification step. Claude will then call out to Codex, pass the context, get verification, and incorporate the feedback before giving you the final answer.
For Everyone Else: Claude Cowork + Chrome Extension
If you’re less technically inclined and using Claude Cowork, install the Chrome extension and tell Claude:
Update the CLAUDE.md so that whenever you do some research, you will verify all of your findings and assumptions through ChatGPT. You will open up ChatGPT through the web interface and paste all of the context necessary for it to provide you with feedback and verification. Then use that answer and feedback to do more research to verify it is actually correct. Make sure to incorporate all of the feedback into the final answer you provide me.
That’s it. Claude will now open a browser tab, paste its findings into ChatGPT, read the response, and use that verification to improve its answer. It sounds wild, but it works.
Bonus: ChatGPT Deep Research Is a Hidden Weapon
I know that OpenAI has its dark side. Yet from an engineering standpoint, their models and the value that they provide for just $20 a month is ridiculously good. I am sure that I’m burning through more than $20 a month in just electricity for how much I use it.
The “Deep Research” feature in ChatGPT is amazing for doing market analysis and finding prospects. It’s pretty close to Perplexity Pro. You can get Claude Cowork to take advantage of it, but you only get 10 of those per month so use them wisely and tell Claude to be patient.
Key Takeaways
- Claude (especially Opus) is brilliant at generating and reasoning, but it tends to move fast and make assumptions; that’s where verification comes in
- Your single most powerful prompt is “Are you sure? Double check everything and don’t make assumptions”
- For Claude Code users: set up your CLAUDE.md to have Claude call Codex CLI for verification (skips the UI)
- For Cowork users: install the Chrome extension and instruct Claude to verify through ChatGPT’s web interface
- ChatGPT’s Deep Research is excellent for market analysis and prospect research (10 per month, use wisely)
- Using multiple AI models together is more reliable than trusting any single model alone
If you want more pro tips like this on a weekly basis, sign up for our newsletter. And if we haven’t actually met yet: I’m Dmitri Sunshine, CEO of Solanasis. We provide fractional CISO services, meaning we make sure organizations aren’t going to trip over themselves due to poor cybersecurity practices, and help them implement AI responsibly. Let’s talk.