The moment that made me want to understand this I've been building FinMentor — a multi-agent financial advisor that runs on Claude. Four agents: a portfolio analyst, a market researcher, a macro economist, and a critic that reviews the others before the final answer goes out. It connects to my IBKR brokerage account. I use it daily. One afternoon I ran a portfolio query — something like "how concentrated am I in tech, and should I be worried?" — and the response came back wrapped in so many caveats it was almost useless. The actual analysis was solid. But it was buried under three paragraphs of "this is not financial advice" and "it's important to consider your personal circumstances." I'd seen this before. I always blamed my system prompts. So I rewrote them. Tighter, more direct, explicit instructions to be concise. Same pattern. I tried a completely different prompt structure. Still there. That's when I stopped blaming my prompts.…