How I built this

A long-running AI chat keeps the record of my product decisions. It turns each piece of work into a written directive: what to do, what's out of scope, and where to stop and report.

A coding agent carries out each directive on my machine. It investigates first, then does the work, then reports back. It doesn't make product decisions.

For harder calls I put the same question to several AI models separately, to pressure-test my thinking. Positioning and the homepage rewrite both went through this.

The final calls are mine, and I deploy. Models challenge a decision; they don't make it. On one badge a model pushed for "One-sided", and the product says "Unconfirmed".

Models aren't users. Nothing in this setup replaces user research, and I don't present it as research.

Built the same way: a written brief, a coding agent, and checkpoints where I review the work. Every number on it is mapped to its source in a file I review.

Each RiskFloor briefing is written by a panel of AI roles working from one evidence file that code assembles: prices, key levels, news and scheduled events. One role argues the upside, another the downside, a third looks for weak or unsupported claims, and three judges (the same model given three different briefs) weigh the result. Code, not the AI, calculates the price levels and combines the judges' views.

AI models proposed parts of this structure. My job was deciding how the panel may behave and what reaches the trader. Three rules I hold it to:

• It describes; it never instructs. The briefing weighs the evidence and doesn't tell anyone what to trade.
• Evidence first, conclusion second. I removed my own large lean badge from the top of the briefing because it read as the answer.
• No number that doesn't mean anything. When testing showed the directional label did no better than chance, it came off the page.