Let AI's Review Each Other's Work

Building my investing engines, I don’t just trust one AI’s first draft. Here’s the process I landed on.

I have one AI draft a working version of what I have in mind, then go back and forth with it until I feel we have a refined, finished product or design. That part’s normal. The part that changed things for me: I hand that draft to a different AI and ask that AI to find problems… issues, weaknesses, and just as important, what’s actually strong about it.

Why does swapping AIs matter? Each AI company is building a genuinely capable “brain,” but built by different teams, trained on overlapping (but probably not identical) data, tuned toward different defaults. Ask one AI to review another’s database design, and swap the order, and you’ll get different complaints back. One tends to worry about edge cases, another about naming and structure, another about performance. Just working with any team in your office, rarely is there one “perfect” answer or design. They just notice different things, the same way two experienced engineers would.

I don’t do this for everything. Effort scales with what’s actually at stake:

• Lower priority (marketing copy, a quick blurb): one AI, not worth burning my tokens.
• Something that carries some weight (a proposal, a client-facing doc): have a second AI review.
• Critical (for me, that’s early-stage planning on designing a database for a SaaS product, or anything highly technical / analytical): three or four AIs review it independently.

For the higher importance items, I bring every reviewer’s opinion back to the first AI, the one that drafted it. I ask that first AI to collapse all the feedback into one edited version for me to review. No point in me reading four docs that 90% say the same thing, just to find the 10% where they diverge. The first AI already has the full context of what was written and why, so that’s the natural place for the opinions to converge before I make the final call.

To keep the token-burn as low as possible, after an AI finishes a draft or a review, I ask that AI to save the response as a downloadable markdown (.md) file. I save that file and upload it to the next AI instead of re-pasting or re-explaining everything in chat. I then ask the follow-on AI’s to review, critique, report on the strengths, weaknesses, recommendations and post the response in a downloadable markdown file also. It’s closer to ‘natural language’ for AI than a PDF or DOCX.

The output is better than any single AI can produce alone just as two minds are always better than one.