The Six-Eyes Principle: How I Work With AI

When I started this project, I thought working with AI meant: I ask for something, the AI does it. One human, one model. That is not how it works, at least not well enough for a system that will soon trade with real money.

What does work is what I call the six-eyes principle here. Three parties look at every decision. Me, and two different roles for AI. Three pairs of eyes.

Who does what

I describe what I want. A wish, a problem, a direction. I decide and I carry the responsibility.

Then there is an AI in the role of sparring partner. It thinks along about the architecture, asks clarifying questions, points out risks, and writes out the instruction. This AI builds nothing. It thinks and it reviews.

And there is Cursor, the tool that actually types the code. Cursor gets a precise instruction and carries it out.

I describe, one AI thinks, Cursor builds. Then the thinking AI reviews what Cursor made. No perspective may be skipped.

Why not just let one AI do everything

This is the core. One AI that both designs the solution and builds the solution is checking its own work. And an AI tends to agree with itself. If the same model picks the approach and executes the approach, it misses exactly the mistakes that sit inside the approach itself.

By separating the roles you get real review. The sparring partner AI did not type the code, so it looks at it fresh. It can say: this is wrong, or this should have been done differently. Just like a second developer reviews a pull request without an emotional attachment to the lines.

An example

While designing the homepage, Cursor came up with a centered, narrow layout. Mobile-first, the default choice. On a desktop it felt too narrow and too long.

Instead of immediately handing over a new instruction, the sparring partner AI first made a few visual sketches. Do we mean the same thing? Only once that was clear did a focused instruction go to Cursor. Cursor built, I reviewed, and after three iterations the right shape was there. Three iterations, not one. But every iteration corrected a real insight, not a matter of taste.

Who decides

Important: the AI advises, I decide. For factor weights, trading rules, capital allocation and the go or no-go for going live, the decision is mine. The sparring partner is allowed to push back hard, that is exactly its job, but the final word is mine. It is my money.

The takeaway

The six-eyes principle is not bureaucracy. It is the recognition that one perspective, even an AI perspective, is too little for work that genuinely costs something when it goes wrong. Separate the thinking from the building, and you get review worthy of the name. A second AI that is not attached to the code catches what the first one misses.

Three pairs of eyes is slower than one. But for a system that trades with real money, that is not a cost. That is the insurance.