Nick Perry explains how to use a "canary" instruction, like requiring the AI to start every response with your name, to detect when a model is breaking down. When the model stops following this tiny directive, it signals the AI is about to hallucinate or make wild assumptions, and you should start a new session.
You must make Claude say your name every single time. And the reason why isn't shocking at all. See, Peter Steinberger, the guy behind Open Claude, talks about these things called canaries. They're like an oil check for AIs. The way the AI pros use them is in their instruction files. They add a tiny instruction that says, "Start every response with my name," which acts as their canary. So now when your AI begins to make stuff up, you can catch it quicker. Because AI doesn't go from perfect to shit overnight. It progressively ignores small instructions until it starts making wild assumptions. Which is why if the canary dies, it's a good chance it's time for a new Session. [FOREIGN LANGUAGE]
