So, there's things you're fighting against when trying to constrain the behavior of the llm.

First, those beginning instructions are being quickly ignored as the longer context changes the probabilities. After every round, it get pushed into whatever context you drive towards. The fix is chopping out that context and providing it before each new round. something like `<rules><question><answer>` -> `<question><answer><rules><question>`.

This would always preface your question with your prefered rules and remove those rules from the end of the context.

The reason why this isn't done is because it poisons the KV cache, and doing that causes the cloud companies to spin up more inference.