Nice! I used to do this by starting a session with a random question or string to introduce some noise, but it stopped working as models became more capable.

I'm curious how you extract decisions from the prompt in your tool, is it another model breaking it down?