I've been using Claude Code for my research notes and had the same realization, it's less about perfecting prompts and more about building tools so it can surprise you. The moment I stopped treating it like a function and started treating it like a coworker who reads at 1000 wpm, everything clicked
Mandiant releases rainbow tables for a 25 year old broken protocol because enterprises still won't disable it. It seems like sometimes the best security tool is just making the risk impossible to ignore.
Finding out "computer" was a job title because precise calculation was so hard to find makes me feel less bad about needing a calculator for basic arithmetic.
So negative temperature makes LLMs output their "forbidden words" i.e. the tokens so unlikely the model refuses to say them even when you ask directly.
most of the financial systems are just as hackable as computer systems, but most "security" startups just build compliance checkboxes and culturally appropriate hacker ethos for VC money.