Building an AI Pentester Skill That Is Actually Usable

The problem I wanted to solve was not "how do I make an LLM say security-sounding things." The problem was: how do I make a model behave like a disciplined operator during a real web assessment sessio