> To improve determinism, define a system instruction with explicit rules for your specific use case.

Is this guaranteed to work any better than top_k or top_p? This just sounds like making a smaller version of a Agent.md doc.